[functorch] Add `functorch_additional_op_db`
PyTorch OpInfo coverage doesn't include nn.functional.* ops. We can use
functorch_additional_op_db as a staging ground where we implement
nn.functional.* OpInfos and then later upstream them.
Why not just upstream them immediately? I can write OpInfos a few times
faster if I don't have to worry about figuring out what the correct
flags are to pass PyTorch tests...
