Add support for torch.ops.aten._to_copy.default

Lower torch.ops.aten._to_copy.default to TOSA CAST op. This resolves
issues around arithmetic operators when using int scalars in unquantized
networks (see new test cases in test_scalars.py).

Note: Parameter 'memory_format' is not supported.

Change-Id: I7a921ca510c5b46f15b5399218f9230ba0f93d88