Onboard ragged sorting kernels from tpu inference to maxtext

[NuojCheng]

Description

This PR onboards two critical kernels: ragged_gather and ragged_gather_reduce.

The key difference lies in their routing and output shapes:

• ragged_gather: Performs simultaneous permutation and fan-out (num_tokens x emb_dim → num_tokens x top_k x emb_dim).
• ragged_gather_reduce: Adds an accumulation step for fan-in (num_tokens x top_k x emb_dim → num_tokens x emb_dim).

During the forward pass, we use ragged_gather for dispatch and ragged_gather_reduce for combine. In the backward pass, these roles are swapped.

FIXES: b/496676734

Tests

We observe performance advantage when ragged sort is enabled, while also notice some regression issue between JAX==0.10.0 and 0.10.1.

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

❌ Patch coverage is 11.77945% with 352 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/maxtext/kernels/ragged/ragged_gather_reduce.py	9.09%	159 Missing and 1 partial ⚠️
src/maxtext/kernels/ragged/ragged_gather.py	9.70%	120 Missing and 1 partial ⚠️
src/maxtext/kernels/ragged/ragged_sort.py	8.45%	65 Missing ⚠️
src/maxtext/layers/moe.py	66.66%	4 Missing and 2 partials ⚠️

📢 Thoughts on this report? Let us know!