backwards of autograd functions

[t-vi]

661 allows us to trace through torch.autograd.Function .apply using the forward, but so far we ignore the backward (and emit a warning). It would be neat to fix this and also trace the backward / use it in a grad_transform or so.

[tfogal]

triage review:

• Currently we can drop some custom autograd handling on the floor
• We should do this properly but needs some proper design or discussion
• Can we run these as a black box through torch instead of understanding them?
• We don't run jit during backward right now... when do we want to acquire backward?
• Users who care enough can reimplement fwd+bwd in their own transform