Muon and explicit Muon + AdamW groups

Select hidden nonempty **complete 2D matrices** explicitly. Embeddings, classifier heads,
biases and normalization parameters should normally use AdamW, even when they are 2D.
All unselected parameters use the existing high-level AdamW implementation, NOT the optional fused-AdamW interface.
The example learning rates are not tuning recommendations.

`Optimizer::step(lr, ...)` uses `lr` for Muon and `lr * adamw_lr_ratio` for AdamW (ratio default 0.015).
`try_step_with_lrs` allows independent schedules. Missing gradients skip both decay and momentum.
`try_step_or_skip(..., true)` skips both groups without updating state.
Metadata is checked for both groups first, but device failures remain asynchronous: this is not a device transaction.
Tied parameters are routed once by ParamId; callers must not represent overlapping weights with unrelated IDs.

## Numerical choices and migration

Existing constructor defaults preserve SGD momentum, tensor-dtype normalization and AsStored scaling.
EMA mode is opt-in: `m = beta*m + (1-beta)*g`, zero initialized; Nesterov uses `(1-beta)*g + beta*m`.
Do not interchange EMA and legacy SGD checkpoint buffers without a deliberate conversion.
The finite quintic Newton-Schulz polynomial is not an exact polar decomposition; an exact identity-matrix assertion is invalid.

Stable normalization first scales by the maximum magnitude before forming the sum of squares.
It is opt-in because rounding changes, and explicitly requires FP32 inputs/state.
It does not protect against overflow in every other stage and does not inspect nonfinite values.
No implicit BF16 cast is performed. Native PyTorch Muon uses BF16 for NS; the FP32 variant is not bit-equivalent.
No automatic FP32 master weights or low-precision model-copy update is provided.

RUDA Linear stores `[input, output]`, so use InputOutput for Original LR scaling when appropriate.
AsStored interprets rows as outputs. MatchRmsAdamW is symmetric under transpose.
Weight decay uses the original, not shape-adjusted, learning rate.
One configuration covers the selected Muon group; mixed logical layouts need separately configured optimizers.

The project's Config macro does not supply serde defaults for new fields.
Old JSON configs must add `"momentum_mode":"Sgd"`, `"stable_normalization":false`,
`"matrix_layout":"AsStored"` to preserve the old choices. Programmatic defaults remain available.
Simple Muon tensor-record layout is unchanged. Mixed records include a version, configuration,
parameter identity/shape/dtype manifest and both optimizer states. Restore the model with original IDs first.
FullPrecisionSettings is needed for precise continuation comparisons. Changed grouping/configuration is rejected.

Unscale and check gradients externally before updating, and synchronize skip decisions across replicas.
This patch does not automatically integrate the prior low-level gradient-guard API.
Implicit `step_multi`, Ruda distributed-marked tensors, FSDP/TP shards, sparse gradients and 4D convolution reshaping
are not implemented. Never orthogonalize arbitrary shards as though they were the full matrix.

## Validation

```bash
cargo run --release --locked -p ruda-optim --example muon-training -- 20
python tools/run_muon_regressions.py --suite oracle
python tools/run_muon_regressions.py --suite reference
python tools/run_muon_regressions.py --suite host
python tools/run_muon_regressions.py --suite build
python tools/run_muon_regressions.py --suite cuda --compiler both