[functorch] convolution_backward batch rule (pytorch/functorch#359)
This is the most ridiculous batching rule we have.
Featuring a guest appearance of efficient zeros tensors.
We should really consider upstreaming einops.rearrange
(https://einops.rocks/api/rearrange/). Many batching rules are straight
up dimension manipulation and if we could specify that with strings we'd
be done 10x faster.
Test Plan:
- run tests
