Speed up default FNet testing speedups.

[Cyber-Machine]

Partially fixes: #866

• Add a serialization test for all classes.
• Mark the saved model testing as "large" for all classes.
• Make the backbone sizes as small as possible while still testing all logic (e.g. num_layers=2).
• Only run a single jit_compile=False test per model class. E.g. test_classifier_fit_no_xla.
• Other minor readability and efficiency improvements (check the original PR carefully).

[mattdangerw]

This looks like it has some legitimate test failures to work through before we can merge. Related to sentencepiece it seems.

[mattdangerw]

Let's fix the tests!

[Cyber-Machine]

@mattdangerw, Will be working towards checking all the tests.

[chenmoneygithub]

Thanks for the PR! One minor comment.

Meanwhile, could you post the time cost change before and after this PR? You can run

pytest keras_nlp/models/f_net 

on master and your branch separately.

[chenmoneygithub]

Let's reduce the vocab size too.

[Cyber-Machine]

Thanks for the PR! One minor comment.

Meanwhile, could you post the time cost change before and after this PR? You can run

pytest keras_nlp/models/f_net 

on master and your branch separately.

Sure, I have performed tests on my FNet-test-speedups branch and master branch, and here are the results.

Before Changes:

• master branch
• Time taken: 140.05s (2mins and 20 sec)
• 55 Tests passed, 32 Tests skipped

After Changes:

• FNet-test-speedups branch
• Time taken: 45.73s
• 40 Tests passed, 38 Tests skipped

[chenmoneygithub]

Thanks!