Add batched GPU NMF support with sklearn-compatible multiplicative-update (MU) and Fast-HALS coordinate-descent (CD) solvers

[JerryIshihara]

Changes

• Add the cnmf.gpunmf package, split into:
  • shared GPU/runtime utilities
  • batched MU solver
  • batched Fast-HALS CD solver
  • fused Triton CUDA CD kernel
• Add --solver mu|cd selection.
• Add GPU engine configuration for device, dtype, TF32, compilation, and replicate batch size.
• Support batched factorization while preserving one spectra output per cNMF replicate.
• Enforce solver/loss compatibility: CD supports Frobenius loss only.
• Preserve the existing CPU/sklearn path as the default.
• Replace the previous nmf_gpu.py with focused modules.