One-dimensional real-FFT FNO blocks

The dense reference and native cuFFT paths share every parameter. Each block stores real and imaginary spectral weights, a pointwise skip matrix, and a bias, in that order. The spectral weights have shape [modes, width, width]; unlike the full-DFT model, they do not store separate negative-frequency parameters.

def Runtime.Autograd.Model.Layers.FNO.rfftBlock (grid width modes : ℕ) (hgrid : 0 < grid) (hwidth : 0 < width) (hmodes : modes ≤ grid / 2 + 1) (path : F.SpectralPath := F.SpectralPath.automatic) (activation : Activation.Kind := Activation.Kind.relu) (spectralRealSeed spectralImagSeed skipWeightSeed : ℕ := 0) : Layer [grid, width] [grid, width]

One-sided spectral block followed by a pointwise skip connection and an activation.

The default ReLU and parameter order agree with Cuda.Fno1dRfftFused. Comparisons must still load the same parameter tensors: independently seeded builders need not draw the same values. Changing only path preserves the function and the checkpoint layout.