[functorch] Fix convolution batch rule in the transpose case (pytorch/functorch#345)
We were making wrong assumptions about where the input_channels /
output_channels were in the weight tensor and where the groups dimension
gets included.
Test Plan:
- run tests
