Support TP-sharded matrices in DistributedMuon

[weifengpy]

Stack from ghstack (oldest at bottom):

• -> Support TP-sharded matrices in DistributedMuon #4051
• Add standalone distributed Muon optimizer #4034

Stacked on #4034.

Summary

• extend Add standalone distributed Muon optimizer #4034 Owned() compute to rank-2 matrices stored on a named 2D mesh with exact Shard(0), Shard(1) placements
• use the exact 1D redistribution mesh supplied by BucketSpec; the generic Add standalone distributed Muon optimizer #4034 runtime derives uneven rectangular storage blocks from each DTensor mesh coordinate
• reconstruct each full matrix on its assigned owner, run Muon once, and return update blocks through the reverse all-to-all
• preserve the original parameter, gradient, and momentum DTensor placements

The extension is intentionally limited to uniquely tiled Shard(0), Shard(1) storage. TP-replicated Shard(0), Replicate() storage and _StridedShard redistribution remain out of scope.

Test plan

• four-GPU parity test with uneven 5x7 and 6x5 matrices on a named 2x2 FSDP/TP mesh
• verifies one forward and one reverse all-to-all for the bucket, matches plain Muon numerically, and preserves parameter/momentum placements (1 passed)