Have the ability to disable __torch_function__ dispatch for torch.nn.functional functions

[jamesr66a]

torch.nn.Module instances and torch. namespace ops represent a pretty good split between high-level (stateful) blocks and low-level (stateless) operations. torch.nn.functional calls seem to just get in the way in the vast majority of cases that I've seen. The fact that these participate in __torch_function__ makes this weird middle-layer of abstraction that cannot be easily skipped. I can't even figure out how to skip them manually in a __torch_function__ handler, we can't redispatch into the functional with a layer of tensor-like objects stripped because then we wouldn't get the calls inside the functional (but not the functional itself). Can we have some way to turn off dispatching to __torch_function__ for torch.nn.functional calls?

(Additionally there's a separate issue where torch.nn.functional calls have this icky TorchScript boolean_dispatch thing

cc @hameerabbasi @rgommers @peterbell10 @albanD @mruberry @jbschlosser

[jamesr66a]

Note I have been able to hack something up by monkey-patching has_torch_function in torch.nn.functional:

        to_patch = ['has_torch_function', 'has_torch_function_unary', 'has_torch_function_variadic']
        try:
            def no(*args, **kwargs):
                return False 
            for name in to_patch:
                locals()[name] = getattr(torch.nn.functional, name)
                setattr(torch.nn.functional, name, no)

           <snip>
        finally:
            for name in to_patch:
                setattr(torch.nn.functional, name, locals()[name])

but this is bad, to say the least