Salvatore Mesoraca

When I tested it, it passed all tests (e.g. test-backend-ops). So it seems to work, but I don't know if there is any corner case in which it would return...

> > Implementing conv_transpose_1d via mulmat+col2im uses more memory but it can be ~80% faster (based on a very simple benchmark I did) than the current approach when the input...

Oh wait. I thought that your > I'm not convinced that the code in this PR is necessarily an upgrade over the code on master was specifically referring to something...

BTW, if the increased memory usage is a concern, instead of re-implementing the existing `conv_transpose_1d`, I could add a new `conv_transpose_1d_gemm` and let users decide for themselves.