Fix: add num_hidden_layers property to T5GemmaConfig and add test for use_cache

[priyankabolem]

This PR fixes a bug in T5GemmaConfig where the configuration did not expose num_hidden_layers, which is required by generation/cache utilities (use_cache=True).
• Added num_hidden_layers property to T5GemmaConfig.
• Ensured fallback to decoder’s layer count if not set explicitly.
• Added a unit test (test_generation_t5gemma.py) to verify generation runs successfully with cache enabled.

Testing
• Added test_generate_use_cache_works_for_t5gemma.
• Verified test passes locally with pytest.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: t5gemma

[priyankabolem]

Local pytest for t5gemma passes. The failing CI jobs seem unrelated. Please advise if changes are needed.

[ydshieh]

Hi @priyankabolem Could you share an example (or a test in the library) that is failing prior to this PR?

[priyankabolem]

Hi @ydshieh, thanks for the feedback!

I ran the test locally before and after the fix to confirm the issue:

• On main (before fix): test_generate_use_cache_works_for_t5gemma fails with
  AttributeError: 'T5GemmaConfig' object has no attribute 'num_hidden_layers'.

• On my branch (after fix): the same test passes successfully.

This shows that the fix (adding num_hidden_layers to T5GemmaConfig) resolves the issue.