Match GPU NMF denominator handling to sklearn

[JerryIshihara]

The previous implementation added eps to every denominator:

W = W * ((Xg @ Ht) / (W @ (H @ Ht) + eps))

scikit-learn only replaces denominators that are exactly zero. Adding eps to valid, nonzero FP64 denominators slightly changed every update, while the same 1e-9 addition generally rounded away in FP32.

The update now leaves nonzero denominators untouched and replaces only exact zeros:

denominator = W @ (H @ Ht)
denominator = denominator.where(denominator != 0, eps)
W = W * ((Xg @ Ht) / denominator)

The same change is applied to the H update and the fixed-H path, and the default replacement value now matches scikit-learn's FP32 epsilon.