Add backward ops for spatial nodes and spiking neurons

[w4ffl35]

Closes #40, closes #41.

Backward ops following the existing _backward conventions in ops.h:

• conv2d/conv1d: transposed grad_input, accumulating grad_weight (flat, same linear order as params->weight) and grad_bias; grad_bias NULL iff no bias, mirroring affine_backward.
• flatten: reshape the flat gradient to the retained input shape.
• avgpool2d/sumpool2d: scatter over each window (count-include-pad divisor for avg).
• scale: grad_input = gy*s plus an opt-in accumulating grad_scale (scale stays a fixed graph constant by default).
• lif/if/cuba: surrogate through the spike output, zero gradient through the hard-reset branch (grad_stored accepted for tape wiring, contributes nothing); membrane recomputed from retained input/prev state. Reset convention documented in ops.h.

Tests: new tests/test_backward.c with hand closed forms, a central-difference sweep of the conv forward ops (stride/pad/dilation/multi-channel), both spike branches pinned by the forward op, and wiring rejections. Debug, Release, and ASan/UBSan ctest runs pass 36/36 locally.