feat(muon): split fused QKV projections

[RangiLyu]

Motivation

Muon should orthogonalize the logical projections inside a fused QKV weight independently.

Treating a fused QKV weight as a single matrix couples unrelated Q, K, and V updates during Newton-Schulz orthogonalization and applies learning-rate scaling based on the combined matrix shape.

This PR uses the get_muon_split_sizes() interface introduced in #2001 to expose the logical row blocks of fused QKV parameters.

Changes

• Add Muon split metadata for GatedDeltaNet in_proj_qkv.weight:
  • Q: key_dim
  • K: key_dim
  • V: value_dim
• Add Muon split metadata for Qwen3-VL Vision qkv.weight:
  • Q: dim
  • K: dim
  • V: dim
• Keep the forward computation and parameter layout unchanged.

The Muon optimizer configuration discovers these mappings through the module-level get_muon_split_sizes() interface and processes each logical projection independently.