Fused kernels

The spline part of a layer can run as a fused kernel that never builds the (batch, in, size) basis tensor. Three primitives close under differentiation, so gradients of any order work (PINNs, gradient penalties):

F_m(x, c, w)[b,o]   = Σ_{i,k} w[b,i] c[o,i,k] φ_k^(m)(x[b,i])
G_m(x, g, w)[o,i,k] = Σ_b     w[b,i] g[b,o]   φ_k^(m)(x[b,i])
P_m(x, g, c)[b,i]   = Σ_{o,k} g[b,o] c[o,i,k] φ_k^(m)(x[b,i])

Polynomial and Fourier derivatives are folded into the coefficients; B-splines and bumps compile the derivative order into the kernel.

platform	kernels	backend="auto" uses them for
PyTorch MPS	Metal	B-splines, bumps (RBF, wavelets, RSWAF)
MLX	Metal	B-splines
PyTorch CUDA	Triton (IEEE float32)	B-splines
JAX GPU	Pallas	never (XLA's plain path is faster)

backend="kernel" forces them (error if the basis or device has none), backend="ops" disables them. Bases with trainable parameters always use plain ops. Numbers in Benchmarks.