Support padding channels to ensure gapping conditions for 2d conv

[asraa]

For nchw_fchw 2-D convolutions, the gapping factor g must satisfy g^2 divides the number of channels. otherwise, the channels have to padded first to support the shuffling operation.

Fall out from #2375

[github-actions]

This issue has 1 outstanding TODOs:

• lib/Transforms/LayoutPropagation/LayoutPropagation.cpp:715: handle padding the output channels to be divisible by gap^2.

This comment was autogenerated by todo-backlinks