core/nn/prelu library
Parametric ReLU: y = max(x, 0) + slope[c] * min(x, 0), with a
learnable per-channel slope. Used by MTCNN's P/R/O nets (all
convs are Conv2d + PReLU + optional pool).
Implemented as a host-side scatter so it works on CPU or GPU input (round-trips through Float32List and returns on the same device).