nx_f32_conv2d_grouped.nx

nx_f32_conv2d_grouped.nx -- software-f32 GROUPED 2D convolution (NCHW), generalizing nx_f32_conv2d with a `groups` parameter. groups=1 is the dense conv; groups=C_in (with C_out a multiple) is DEPTHWISE -- the op a MobileNet/EfficientNet-class pose backbone is almost entirely built from, and the single biggest gap the reconnaissance found (nx_f32_conv2d has no groups; the graph-IR has a DEPTHWISE kind with no kernel behind it). Input C_in and output C_out are each split into `groups` contiguous groups; output channel co (in group g=co/(C_out/groups)) convolves ONLY the input channels of group g. Weight layout matches PyTorch grouped conv: [C_out, C_in/groups, KH, KW]. Composes ONLY nx_f32_mul / nx_f32_add. license_tier: ORIGINAL