Minimal KAN-based approximator for y = sin(x) on 0, pi/2.
Uses a single KAN layer with a small grid and identity activation; basis/mixing tensors are learnable via the usual parameter updates.