nx_f32_conv_transpose2d.nx

nx_f32_conv_transpose2d.nx -- software-f32 2D TRANSPOSED convolution / deconv (NCHW), a STANDARD vision op the f32 tower lacked (only a 1-D conv_transpose in nx_vocops). It is the learned-upsampling head of a SimpleBaseline/ ViTPose pose net (3 stride-2 deconvs 8x-upsample the low-res features into full heatmaps), and is broadly used by segmentation/generative decoders. Mechanically it is the SCATTER dual of a conv: each input pixel scatter-adds its value * kernel into a strided output window. Weight layout is PyTorch ConvTranspose2d: [C_in, C_out, KH, KW]. OH = (H-1)*stride - 2*pad + KH + out_pad. Composes ONLY nx_f32_mul / nx_f32_add. license_tier: ORIGINAL