Now in both inits and spectral norm, we assume that dim 0 of weight is the output feature dim. However, this is incorrect for ConvTranspose.

We need to figure out how to properly handle those.

cc @albanD @mruberry