In VariableType.cpp, #16305 adds checks to make sure we are still using the same input TensorImpl and Storage after we call the non-autograd in-place function. However, it is still possible that the in-place function changes the data_ptr of the input tensors without changing TensorImpl or Storage (e.g. by calling tensor.resize_({bigger_than_original}). If the input tensors have 0-size, then we expect this behavior to happen and all is good; if the input tensors have non-zero-size and is smaller than needed, we want to tell the user to resize their input tensors to the correct size before passing them into the VariableType function, to prevent tensor.resize_({size_needed}) from allocating a new data_ptr and to make sure the buffer they pass in would be the buffer that contains the returned data.

cc @ezyang @albanD @zou3519 @gqchen @pearu @nikitaved @soulitzer @lezcano @Varal7