Add standalone distributed Muon optimizer

[weifengpy]

Stack from ghstack (oldest at bottom):

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

Summary

• add a standalone CUDA-only DistributedMuon that owns momentum and Newton-Schulz math without depending on torch.optim.Muon
• add components/distributed_optimizers/ with a generic bucketed redistribution runtime, Muon state/math, and parameter-view preparation
• make redistribution topology a runtime contract: every BucketSpec carries its exact 1D DeviceMesh, owner ranks are mesh-local, and each bucket may use a different participant set
• let the generic runtime derive process groups and DTensor storage blocks, build logical routes and physical schedules, fuse bucket packing, and perform storage-to-compute redistribution and writeback
• require forward and reverse routes to be in-bounds, exhaustive, non-overlapping logical block partitions; every compute destination must receive the complete logical tensor before it may source writeback
• represent replicated or reduced physical copies through participant tuples on one logical block rather than overlapping block entries
• resolve static DSV3 BucketConfig(mesh_axis=...) entries against named DTensor storage meshes after model parallelization
• lower each active Shard-to-Owned bucket into one forward and one reverse packed all-to-all while preserving zero-sized storage shards
• support Owned compute from an exact 1D Shard on either matrix dimension; Support TP-sharded matrices in DistributedMuon #4051 adds named 2D Shard(0)/Shard(1) storage
• expose typed MuonComputeSharding(view_before_placement=..., placement=...); Muon supplies compute requirements and math but no process groups, storage blocks, or collectives
• preserve DTensor parameter and momentum storage placements and checkpoint behavior
• verify before each step that locally computed cached views still share the current DTensor storage, failing before mutation if storage was replaced
• add a DSV3-16B FSDP2=8, EP=4, spmd_types recipe; projection matrices use Muon while wo, embeddings, lm_head, and unsupported parameters remain in foreach AdamW

This is a standalone PR based on main; it has no dependency on the earlier FlexShard/MuonAdapter stack or PyTorch Muon changes.

Test plan

• PYTHONPATH=/tmp/pytorch-optimizer-step-test:$PWD python -m pytest tests/unit_tests/test_bucketed_optimizer_redistribution.py tests/unit_tests/test_muon_parameter_prep.py tests/unit_tests/test_deepseek_v3_distributed_muon_config.py -q (13 passed, 11 subtests passed)
• focused cached-view replacement and numerical/checkpoint tests (2 passed on two GPUs)
• focused uneven Shard(1) parity/writeback test (1 passed on two GPUs)
• focused distinct-bucket-mesh test (1 passed on four GPUs)
• eight-B200 DSV3-16B smoke: FSDP2=8, EP=4, TP=PP=CP=1, spmd_types, local batch size 1, sequence length 128; loss 12.05116, grad norm 8.4805, memory 24.14 GiB per GPU