Implement Winograd as a replacement for ggml_conv_2d

[amblamps]

This repo seems to improve sdcpp performance by quite a bit.
https://github.com/SealAILab/stable-diffusion-cpp

[amblamps]

only improves performance for CPU according to this comment
ggml-org/ggml#971 (comment)