MaxPoolNdBackward

Autograd node for N-dimensional max pooling.

Each output element is the maximum of a $K_0 \times \cdots \times K_{N-1}$ window of the (optionally zero-padded) input: $y[b, c, o] = \max_{0 \le k < K} x[b, c, s\,o + k]$ (extended to $N$ axes). The forward pass also produces a tensor of argmax indices into the flat padded-input buffer; these are saved as saved_argmax_ and consumed by the backward pass to route each gradient back to exactly one input element.