Add GPU fixed-H consensus refit

[JerryIshihara]

Summary

Adds GPU support for the fixed-H NMF refit path used during cNMF consensus.

Changes

• Adds a fixed-H multiplicative-update path for update_H=False
• Keeps fixed spectra H unchanged and updates only usages W
• Adds validation for fixed H shape, finiteness, and non-negativity
• Splits normal NMF and fixed-H refit paths into separate helpers for readability
• Allows --engine gpu for both factorize and consensus
• Keeps the standalone GPU kernel API unchanged