feat(cutedsl): add CuTe-DSL Newton-Schulz operator and LigerMuon optimizer

[piyushumate]

Summary

This PR introduces a CuTe-DSL accelerated 5th-order Newton-Schulz operator and the LigerMuon transformer optimizer to Liger-Kernel.

Muon (MomentUm Orthogonalized by Newton-schulz) has demonstrated significant sample efficiency gains over AdamW for pre-training and fine-tuning LLMs by applying orthogonalized momentum updates to 2D weight matrices:

X_{k+1} = 0.5 * X_k * (3I - X_k^T * X_k)

In eager PyTorch, the 5-step iterative polar decomposition incurs repeated un-fused GEMM launches and high HBM memory traffic across dozens of layers. This PR provides:

1. cutedsl_newton_schulz_forward in liger_kernel.ops.cutedsl.ops.newton_schulz with automated PyTorch fallback.
2. liger_newton_schulz operator exposed in liger_kernel.ops.newton_schulz.
3. LigerMuon optimizer in liger_kernel.transformers.muon, routing 2D weights to accelerated Newton-Schulz and 1D parameters (biases, layer norms) to AdamW.

Testing Done

• Hardware Type: NVIDIA H100 SXM5 80GB (tested via Modal E2E)
• All 11 tests in test/cutedsl/test_newton_schulz.py passed on H100 (numerical parity across shapes/dtypes, singular value bounds, and loss convergence).
• ruff check and ruff format passed cleanly.