nx_f32_bn_fold.nx

nx_f32_bn_fold.nx -- fold inference BatchNorm into the PRECEDING conv's weights + bias (offline), so a ResNet/ HRNet-class pose net (which uses BN, not GroupNorm) loads straight into the existing dense/grouped f32 conv with NO runtime BN kernel. For a conv output x = W*in + b followed by BN y = gamma*(x-mean)/sqrt(var+eps) + beta, the fold is exact per output channel: s = gamma/sqrt(var+eps); W' = W*s; b' = (b-mean)*s + beta. Then foldedconv(in) == BN(conv(in)) identically. Composes nx_f32_mul/add/sub/sqrt/div. This is the standard "BN-fold" every inference runtime does; it means the porter never needs a live BatchNorm op. license_tier: ORIGINAL