fix(muon): avoid EP/FSDP AGRS OOM

[jayhenry]

Summary

• Expose MuonConfig.remainder_strategy="pad_all2all" as a temporary workaround for remainder batches when all-to-all is available.
• Fix the global shard dimension passed to AGRS and all-to-all when EP and FSDP shard the same tensor dimension.
• Add a safe mechanism-level reproduction, an EP+FSDP regression test, and update xtuner_muon_oom.md.

Root cause

For GLM-5.2 with 512 ranks and EP8, the expert FSDP group has 64 ranks. The 76 same-shaped fused_w2 parameters form a remainder batch of 12. Before this fix, AGRS used the full DTensor dimension (256 * 6144) instead of the EP-local dimension (256 * 6144) / 8, producing a 72 GiB buffer and a second 72 GiB flatten copy.

Validation

• PYTHONPATH=. python -m pytest tests/optim/test_muon.py -q — 12 passed.
• CPU and 4-rank CUDA runs of .dev_scripts/repro_muon_agrs_ep_oom.py completed successfully under the local GPU lock.
• The new regression test fails on the parent revision (Newton-Schulz sees 12 rows) and passes with this change (6 rows).

The two commits are intentionally separate: the first is the temporary configuration workaround, and the second is the permanent dimension fix.