nx_f32_conv2d_backward_fast.nx

nx_f32_conv2d_backward_fast.nx -- FAST conv2d backward via im2col + the fork-parallel nx_f32_matmul_t, the companion to nx_f32_conv2d_fast so a full TRAINING STEP is fast (the serial backward is the bottleneck ~2/3 of a pose-net step). Formulation (N=1): colT[K,M] = transposed im2col of input; dWeight[C_out,K] = matmul_t(dOut[C_out,M], colT[K,M]); dCol[M,K] = matmul_t(doutT[M,C_out], weightT[K,C_out]); dInput = col2im(dCol); dBias[co]=sum_p dOut. Bit-exact vs nx_f32_conv2d_backward on integer values. ⚠N=1 (dWeight is assigned by matmul_t, not accumulated). license_tier: ORIGINAL