This pr adds an api torch_xla._XLAC._set_buffer_donation which can be used to mark the buffer associated with current tensor to be donated in the next execution. This api is currently only enabled for dynamo(torch.compile) use case, ti is a no-op in LazyTensor world(since LTC already has the auto aliasing based on inplace op and dynamo's functionization pass remove all inplace ops).

Example usage should be

  def dummy_inplace_add(self, input):
    input += 1
    return

  def test_manual_buffer_donation(self):
    device = xm.xla_device()
    input = torch.randn(5, 5).to(device)
    dummy_inplace_add_compiled = torch.compile(
        self.dummy_inplace_add, backend='openxla')

    met.clear_all()
    # input is a device_data, we should be able to set the buffer donation field.
    self.assertTrue(torch_xla._XLAC._set_buffer_donation(input, True))
    # make sure buffer donation setting is correctly updated
    self.assertTrue(torch_xla._XLAC._get_buffer_donation(input))
	
	for _ in range(100):
      # You don't need to keep calling this function if you function does not cause dynamo recompilation.
      # check below for the reason.
      self.assertTrue(torch_xla._XLAC._set_buffer_donation(input, True))
      dummy_inplace_add_compiled(input)

Please note a couple things

1. _set_buffer_donation is called on a tensor, but being applied to the buffer associated with this tensor. If the tensor you passed in is not a buffer(for example an intermediate tensor that has not been evulated) this api will be no-op and return false.
2. Buffer aliasing is being set up during compilation time. Torch Dynamo also does not track this field, so even if you change _set_buffer_donation after first execution(torch.compile compilation triggered at first execution of the compiled function), aliasing will not change.