TORCH_FUNCTION for Proxy

[Essoz]

This PR resolves dispatch-related compatibility issues caused by wrapping torch objects using Proxy.

The PR resolves it by adding a __torch_function__ implementation to the Proxy definition which automatically invokes any API to be dispatched with the original objects instead of the proxied ones.

After this PR, we are able to instrument the word_language_model input in the pytorch example repo. Previously, instrumentation it leads to crashes as torch cannot find kernel implementations for the proxied tensors.

[Essoz]

Tested the whether this PR allow us to get rid of all the unproxy wrapped applied to builtin APIs to avoid passing Proxy into backend which leads to segmentation fault.

This PR allowed us to pass through the forward phase without any problems. But we crashed at loss computation at F.nll_loss in the MNIST pipeline. Problem seems to be inside the handle_torch_function inside the nll_loss implementation.

[Essoz]

did a little bit deeper search. The code crashes at l1627 in torch.overrides.py.

[Essoz]

Ok the problem is caused because torch_func_method is the regular pytorch Tensor __torch_function__ instead of our own unproxying torch function.

Note how type(args[0]) is a proxy and is being passed to C.

[Essoz]

renaming Proxy._unwrapping__torch_function__ to Proxy.__torch_function__ solved the problem:

torch sometimes rely on hasattr(cls, '__torch_function__') to determine whether torch function needs to be handled for a type and this hasattr isn't controlled by us.