Base64 encoded data-uris can contain newlines, make sure we handle
them correctly.

Content Type errors were incorrectly being set as an array on the base
attribute rather than being set individually as strings.

Copying attribute validation failures up to the root object breaks the
validate_attachment_presence_matcher when other paperclip validations
have failed.  This fixes this bug.

Also refactored to make the strategy hierarchy clearer

and memoize type_from_file_command

You can either specifiy :restricted_characters or you can define your
own :filename_cleaner object. This object needs to respond to #call and
takes the filename that will be cleaned. It should return the cleaned
filenme.

Return the default path when the attachment has no remote file

Using the ActiveRecord validations is smart in order to lessen load on
the project to develop our own validators, but it's problematic in that
calling `valid?` on the record fires off all the other validations --
including those on fields which may not be set yet because of
mass-assignment. This commit will "pre-validate" the attachment's fields
so that it doesn't process an invalid attachment, but it does so by
running its validations manually on assignment.

Refactor Paperclip::Attachment to eliminate code duplication between #extra_options_for and #extra_source_file_options_for

This commit makes this method move by default instead of copying, which
should make it perform significantly better with large files. It falls
back to the old semantics if the mv fails.

This propagates attachment size and content_type validations to the main
attachment attribute, to make them behave the same as presence
validations. This allows paperclip to play more nicely with simple_form
and other form abstraction gems.

This commit includes a behaviorial change. If you emit all errors on a
model instance you will see some failed validations twice: once for the
base attachment attribute and again for the failed presence,
content_type or size attachment attributes.

There was a spec that was failing because we removed .attachment_definitions
from models, I expected this to fix it, but apparently it hasn't.