This enables the deferral of the resolution of the serializer class to prevent NameError exceptions due to reference cycles between serializer classes.
An `_id` suffix will be appended to the name of HasOne associations, while `_ids` will be appended to the singularized name of HasMany associations. Association keys can still be overridden with the `key` option.
Furthermore, objects embedded in the root are now by default named according to their serializer, instead of the key used for their associations.
To achieve this, we make the following change when sideloading: Instead
of serializing associations and discarding duplicate *hashes*, we
memorize the *objects* (records) that we have already serialized, and
only serialize those that are new.
This change is mostly transparent, and brings down serialization time
from 3.1 seconds to 1.0 seconds on my set of sample data.
There is one change in the behavior: If you sideload the same object
multiple times, and it yields different hashes, like so:
embed :ids, include: true
has_many :comments
has_many :recent_comments, root: comments, serializer: CommentShortSerializer
then previously, it would be included multiple times, whereas now, the
first hash wins. (I haven't actually tested this.) I don't know that
either option is preferable. It's not covered by the test suite, and I
think it's an edge case that is OK to ignore entirely.
include/exclude functionality.
This process involves building a few lower-level
primitives so not as much logic is hardcoded in
monolithic helper methods.
