Remove API export decorator from base classes

[kanpuriyanawab]

Fixes #823

Needs @mattdangerw 's confirmation, keeping this comment in mind.

[mattdangerw]

Actually Sampler and Tokenizer we do want to export right now, as we want library users to be able to write subclasses for these.

This looks good for Backbone, Task, and Preprocessor!

[kanpuriyanawab]

Hey @mattdangerw, changes us are up, and PR is RTM !
btw could you quickly summarize how keras_nlp_export differs from register_keras_serializable in terms of functioning, (apart from export logic).
Thanks !

[mattdangerw]

Sure! keras_nlp_export just does register_keras_serializable under the hood. https://github.com/keras-team/keras-nlp/blob/master/keras_nlp/api_export.py

Generally it is good practice to make sure every object in the library is registered as serializable if it (or a parent class) has a from_config(). So that is basically all layers, tokenizers, models, metrics, samplers, etc. Whether or not they are part of the API.

keras_nlp_export means we are adding this to the API of the library. Everything with an export annotation will be under our API contract and documented on keras.io. We have things (like the WhisperEncoder we just added) that we don't want to export and are essentially private details of our library. This includes Task and Backbone classes for now. We don't want people writing their own subclasses and filing bugs if we break compatibility on these classes.

So basically all "keras objects" should have either an register call or an export call. Register if its private functionality, export if its public.

[mattdangerw]

LGTM! Will merge as soon as CI is green.