[torch_xla2] Wire `torch_xla2.compile`d function with torch `AutogradFunction`

[qihqi]

🚀 Feature

Currently if we wrap with model with torch_xla2.compile and want to train the model using the traditional torch training loop similar to https://github.com/pytorch/xla/blob/master/experimental/torch_xla2/examples/basic_training.py

You would notice that it doesn't work.

The reason is because the compile wrapper JittableModule will eventuall call a jax.jitd callable, and torch doesn't know how to compute gradient of that callable.

The solution is to create a torch.autograd.Function subclass on the fly, with backward defined to call jax.vjp similar to this tutorial: https://pytorch.org/tutorials/beginner/examples_autograd/two_layer_net_custom_function.html

The result would be that wrapping a model with torch_xla2.compile it is still trainable.

Motivation

Having the forward and backward compiled with jax jit is faster to run.