[DistMuon] Simplify layout planning and matrix views

[weifengpy]

Stack from ghstack (oldest at bottom):

• [DistMuon] variable-sized BlockShard to shard inside a q/kv head #4799
• -> [DistMuon] Simplify layout planning and matrix views #4809

DistMuon carries overlapping prepared-layout records and separate paths for
native matrices and row-concatenated matrix batches. Resolve layouts once and
build explicit matrix views after local compute ownership is known. Dense
Muon helpers consume ordinary tensors and views; optimizer orchestration
handles sharding and redistribution.

Represent each view with shape, element strides, and a relative offset. Store
the final views by parameter name and rebuild them after checkpoint loading.

This keeps parameter normalization, the scalar BlockShard API, Kimi layouts,
ownership rules, Newton-Schulz batching, and adjusted-LR application order
unchanged. Variable block sizes and per-matrix scaling are introduced by the
next PR in the stack.

Validation:

• CPU FlexShard: 8 existing tests passed.
• GPU DistMuon: 2 tests passed with the original numerical assertions intact.
• 160 native/uniform matrix cases produced exactly the same directions and
  final updates as the original implementation across FP32/BF16 and LR modes.
• All 72 builder/direct-constructor input cases match the original base,
  including parameter-name types, generator handling, and group copying.
• pre-commit run --all-files passed.
• A 30-step Kimi debug run with FSDP=8/EP=8, --debug.seed=42, and
  --debug.deterministic matched the original base 6c2dadb at every loss,
  grad_norm, and learning-rate value. Both runs used PyTorch git1807a24.
  Maximum difference was zero; final loss was 2.655120611190796 and final
  grad_norm was 1.0288628339767456.