[Torch] Support bincount and scatter_add ops

[masahi]

Add numpy style bincount op converter by combinations of existing ops.
https://pytorch.org/docs/stable/generated/torch.bincount.html

This is a part of supporting hummingbird project. Also add a scatter_add op conversion, as requested.

I hope performance would not be too bad, so that we don't have to implement a native bincount op.

please review @siju-samuel @t-vi
cc @interesaaat

[t-vi]

I think the type for bincount would be good to improve, otherwise LGTM. Thank you @masahi

[t-vi]

Thank you @masahi.

[tqchen]

Thanks @masahi for contributing! Thanks @t-vi for reviewing!

[apivovarov]

Does the following code in _bincount() require relay.vm.compile()? Because _op.zeros() is called for a shape which is not known at the compilation

data = inputs[0]
maximum = _op.max(data)
dim = maximum + _expr.const(1, dtype="int64")
counts = _op.zeros(_op.reshape(dim, [1]), out_dtype)

[masahi]

Yes, bincount requires VM compile.

[apivovarov]

I'm working on adding TF bincount support. I found thattf.bincount internally uses tf.shape. But looks like tf.shape is not compatible with relay.vm. Instead of determining the shape at the runtime vm.compiler just creates new parameter which should be provided by user when calling vm.run(). https://discuss.tvm.apache.org/t/tf-shape-and-relay-vm-strange-result/13505