Tensor types that have been opted in to the swap_tensors path.

The following table lists tensor types that have been opted in by default to the swap_tensors path in either _apply or load_state_dict. All other types have not been opted in by default and will only use swap_tensors if torch.__future__.set_swap_module_params_on_conversion(True)

This PR adds the escape hatches mentioned in the above table to torch.__future__

• _{get/set}_swap_overwrite_escape_hatch: which allows tensor types that assigned new params in nn.Module._apply that were opted in to swap_tensors (right now just src/dest on xla/meta to fall back to this behavior
• _{get/set}_swap_load_state_dict_escape_hatch: which allows tensor types that were opted in to swap_tensors path in nn.Module.load_state_dict (for now just to fall back to old behavior

This allows backward compatibility as it provides an option to get the old behavior. In the future, we will need one more escape hatch for tensors that used tensor.data = in nn.Module._apply

Can we let these be private for now?

Stack from ghstack (oldest at bottom):

• -> Add private escape hatches to fall back to pre-swap tensors behavior #126984
• Default traceable subclasses to use swap_tensors path for load_state_dict #126788
• Default meta device to use swap_tensors in nn.Module._apply (.to_empty and .to('meta')) #126819
• Default XLA to use swap_tensors path in nn.Module._apply #126814
• Relax use_count constraints for swap_tensors when AccumulateGrad holds a reference #127313