Extend device data node binding API to not clone specified input tensors

[rpsilva-aws]

In this PR, we extend the _get_tensors_xla_device_data_node binding API to return the same tensor values for a given set of specified unmutated tensor inputs. It currently returns a list of tensors that capture the XLATensor values of the graph inputs. However, these tensors end up creating new ATen tensors, which are effectively clones of the original tensors.

xla/torch_xla/csrc/init_python_bindings.cpp

Line 2919 in c4b45a9

torch_xla::XLATensor::Create(backend_data));

This makes it so that these are not eligible to be aliased at the step barrier. Instead, we extend the API to allow users to specify the list of known input tensors, such that if graph input matches one of those inputs, then the same ATen tensor is returned back to the user as instead of returning a clone.

>>> t4 = torch.tensor(50).to(device)
>>> t5 = t4 + 10
>>> t6 = t5 * 20
>>> [torch_xla._XLAC._xla_get_tensor_id(x) for x in [t4, t5, t6]]
[8, 10, 12]
>>> results_with_input = torch_xla._XLAC._get_tensors_xla_device_data_node([t4, t5, t6], [t4])
>>> [torch_xla._XLAC._xla_get_tensor_id(x) for x in results_with_input[1]]
[8, 18, 19]
>>> results_without_input = torch_xla._XLAC._get_tensors_xla_device_data_node([t4, t5, t6], [])
>>> [torch_xla._XLAC._xla_get_tensor_id(x) for x in results_without_input[1]]
[20, 21, 22]

The input tensors are kept as optional, ensuring that the API change is backwards compatible.

cc: @mcuiaws

[rpsilva-aws]

@tengyifei @qihqi PTAL, as this will help us move away from explicit aliasings for the gradient accumulation (needed for 2.7.1), and with some internal nightly experimental workstreams.

[tengyifei]

Is the argument order swapped? Should input_tensors come first?