Currently paperclip fails with `NoMethodError` on `nil` when file is not
present. Since `nil` is proper response from fog on missing files, and
this method already returns `false` if any error happen (see line 178),
returning `false` here would be properly handled on client code.

[closes #2173]

Adds failing spec to command line dup check

[fixes #1670]
[fixes #2169]

`Paperclip::Storage::Fog#host_name_for_directory` assumes a String-like
object is set and doesn't check if `@options[:fog_directory]` is
callable, while `Paperclip::Storage::Fog#directory` does. This extracts
a new method with the condition and refactors the other two methods to
call it.

Fixes #2018, #2093.

This configuration constant was useful for supporting different versions
of AWS SDK. We don't any longer, and we can set the value directly.

[fixes #2148]

We don't support AWS SDK v1 any longer.

Drops Rails 3 branches.

[fixes #2101]

Rails 3 is not supported anymore. We can safely remove mentions and
snippets for it.

Drop unused variable definition (`options`).

While working on this branch, Kito found that this test fails due to:

https://github.com/rails/rails/commit/6ec8ba16d85d5feaccb993c9756c1edcbbf0ba13#diff-fdcf8b65b5fb954372c6fe1ddf284c78R76

We are not yet sure if it's a bug in paperclip or in Rails itself. With
current `ActiveModel::Errors` implementation the following happens:

```
record.errors # => @messages = {}
record.errors.include?(:foo) # => false
record.errors # => @messages = { :foo => [] }
```

Which bit us in:
https://github.com/thoughtbot/paperclip/blob/69f18375333234b6f395300266e2612936bd242e/lib/paperclip/validators/attachment_file_name_validator.rb#L23

Another related Rails commit:
https://github.com/rails/rails/commit/b97035df64f5b2f912425c4a7fcb6e6bb3ddab8d

I worked around the issue by changing what we assert in this spec. I am
still not sure that this is a bug in current Rails master.

cc @kitop for review

Ref: https://github.com/rails/rails/commit/481e49c64f790e46f4aff3ed539ed227d2eb46cb

`silence_stream` was deprecated in Rails 4.2 and removed in Rails 5
because it was not thread safe. Paperclip is not running the tests in
threaded mode, it's safe to add the method back in a test support
module.

`bundle exec rails -v` was taking seconds every time, making the test
suite run very slow. This changes the code to compare the version
against `ActiveRecord::VERSION`, which is going to be the same one as
`rails -v`, making it much, much faster.

Simplifies condition: we no longer support Rails < 4.2, so we test for
`>= "5.0"` or fallback to `4.2` branch.

- Also bump Travis Ruby 2.2.x version to 2.2.4

We found that uploading large files to S3 would result in a socket error
("connection reset by peer") occasionally and lately much more
consistently. In researching this I saw that many people got this error
when uploading too large of a file without multipart chunking. I would
have assumed fog did this automatically but the default chunk size may
be too high. In order to address this I wanted to drop the chunk size to
100MB.

Rather than hard-code this I opted to expose a `fog_option`
configuration option that lets me pass any additional options I want to
the fog's `#create` call. This is similar to the `fog_attributes` option
implemented in CarrierWave which [addresses the same
problem](http://stackoverflow.com/a/11867978/201911).

We've been running this now for a week in production and it seems to
resolve the issue.

https://github.com/thoughtbot/paperclip/pull/2135

The MD5 sum of an attachment is only necessary if the model has a corresponding
_fingerprint column. If this column is absent, there is no need to calculate the
MD5, which can be an expensive operation for large files.

Accomplish this by deferring the fingerprint calculation using a block. If the
_fingerprint column is absent, the block is never called, and the calculation
is avoided.

this commit primarily uses frozen strings to reduce object creation during interpolation.
the :basename method now uses File.basename(file, ".*") rather than a Regexp. basename may be called multiple times.

by using the attachment name (symbol) and Class as keys we reduce the number of Strings created before hitting the cache

* Regression fix
* Add specs for intermediate_files var in Paperclip::Attachment

[fixes #1908]

  $ rspec spec/paperclip/integration_spec.rb --profile 1
  Testing against version 4.1.7
  ..................................

  Finished in 15.19 seconds (files took 0.60155 seconds to load)
  34 examples, 0 failures

  Top 1 slowest examples (9.54 seconds, 62.8% of total time):
    Paperclip Many models at once does not exceed the open file limit
      9.54 seconds ./spec/paperclip/integration_spec.rb:18

to

  $ rspec spec/paperclip/integration_spec.rb --profile 1
  Testing against version 4.1.7
  ..................................

  Finished in 10.95 seconds (files took 0.59585 seconds to load)
  34 examples, 0 failures

  Top 1 slowest examples (5.33 seconds, 48.7% of total time):
    Paperclip Many models at once does not exceed the open file limit
      5.33 seconds ./spec/paperclip/integration_spec.rb:18

Updates gemfiles to Ruby 2 hash syntax.

[closes #1700]

Both v1 and v2 of the AWS Ruby SDK accept the string literal AES256, no need to differentiate. (thank you, @trevorrowe)

According to the ImageMagick documentation the '@' flag is a special case that
can be used in conjunction with '>' to prevent IM from scaling an image up.
The '<' flag is ignored when using '@'.

http://www.imagemagick.org/Usage/resize/#pixel

This fixes the issue I reported in #1851

* update aruba methods for in_current_dir and check_file_presence
* switch rspec syntax from should to expect
* configure raise_in_transactional_callbacks if AR >= 4.2
* remove paperclip itself from Appraisal
This also updates Travis to use the default bundler (to enable caching) and to build primary rubies first

It should never be the case that `url` returns `nil`. If there is no url
that can be reasonable achieved, as is the case when the record the
attachment is attached to is unpersisted, then `url` should return the
default url.

For some reason the test in this PR would make all other ones blow up.
For reasons not fully understood at the time of writing, the alternate
spec I'd written doesn't. So, we'll replace that test with this one and
it'll be green.

Thanks to MORI Shingo of DeNA Co., Ltd. for reporting this.

There is an issue where if an HTML file is uploaded with a .html
extension, but the content type is listed as being `image/jpeg`, this
will bypass a validation checking for images. But it will also pass the
spoof check, because a file named .html and containing actual HTML
passes the spoof check.

This change makes it so that we also check the supplied content type. So
even if the file contains HTML and ends with .html, it doesn't match the
content type of `image/jpeg` and so it fails.

consistent with public URLs, this also changes the default behavior
for expiring URLs to use the https scheme rather than http.