[Functionalization] Harden XLA_DISABLE_FUNCTIONALIZATION
Summary:
This pull request hardens XLA_DISABLE_FUNCTIONALIZATION by:
1. guard every at::functionalization::functionalize_aten_op calls which
will implicitly wrap the tensors into functional wrappers.
2. use at::native counterpart to replace the above call when the op is
really needed when functionalization is disabled.
Test Plan:
PJRT_DEVICE=CPU XLA_DISABLE_FUNCTIONALIZATION=1 python test/test_operations.py -v
