[Enhance] Muon: unify comm interface & reduce peak memory

[nil0x9]

This PR makes the following changes regarding the current Muon implementation.

• Replace multiple mutually exclusive process group params (process_group, subgroup_process_group, world_size, device_rank, subgroup_size, subgroup_rank, use_agrs) with a single process_group +
  explicit comm_strategy: Literal["agrs", "subgroup_allgather", "all_to_all", "replicated"]
• Reduce peak memory by extracting orthogonalization logic into sub-functions (function scope clears refcount on intermediate tensors) and eagerly deleting temporaries
• Refactor existing tests from test_muon_compile.py into test_muon.py, add independent unit tests for all 4 communication strategies