Add 2d fsdp custom mesh and deprecate previous usages

[NuojCheng]

Description

This PR introduces 2d fsdp custom mesh and rule. Unlike deepseek3-671b-2dfsdp, the new rule does not use fsdp_transpose and expert as TP in attention part.

This PR deprecates config yml deepseek3-671b-2dfsdp.yml and flag use_2dfsdp_sharding. In future, use custom_mesh_and_rule=2d-fsdp for the same functionality. Sharding dump tests are added for protection.

Tests

Comparing before and after change, the sharding debugging info are almost the same, and real performance gets improved by the new mesh and rule. HBM usage gets slightly increased mainly because attention weights are replicated along fsdp_transpose.

• Compilation test: https://diff.googleplex.com/#key=5R6Levjze5hB
• Real train test (v5p-8): https://diff.googleplex.com/#key=4alI63fDyj24

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 50.00000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...t/integration/vllm/maxtext_vllm_adapter/adapter.py	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!