nx_f32_conv2d_backward.nx

nx_f32_conv2d_backward.nx -- the GRADIENT of the software-f32 2D convolution: the backward of nx_f32_conv2d_forward. From dL/dOut it computes dL/dInput, dL/dWeight, dL/dBias. This is the ONE training primitive the sovereign ML substrate lacked for CONV nets -- the autograd/Adam/cross-entropy tower (nx_nofloat_autograd, nx_f32_adam, nx_ta_transformer_gradcheck) was matmul/attention-only -- so it is the first brick of TRAINING our own sovereign pose net (operator: "build 2"). Same NCHW layout + stride/pad index math as the forward, exact-mirrored: every (n,co,oh,ow,ci,kh,kw) that contributed input[in_idx]*weight[wt_idx] to out now scatter-adds dout*input into dweight, dout*weight into dinput, and dout into dbias. Composes ONLY nx_f32_mul / nx_f32_add (no libm). Gradient-CHECKED by nx_f32_conv2d_backward_gate (exact integer-f32 gradients, no tolerance). license_tier: ORIGINAL