Fix: Added sharding Constraints for MTP block (b/481469708)

[parambole]

Description

The fix involves adding a single, targeted sharding annotation to target_token_embedding within the MultiTokenPredictionLayer.__call__ method using sharding.maybe_shard_with_logical. This ensures that sharding is correctly propagated through the MTP layer's internal operations, including the projection and the subsequent transformer layer.

FIXES: b/481469708

Tests

Verified with train_compile.py on the TPU backend and validated on ironwood run, where the model now trains without erros.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[suexu1025]

Thanks @parambole

[RissyRan]

LGTM, just one comment.