[LTC] Teach code-gen about composite ops with wrapped numbers (#66912)
Summary:
This patch adds support to code-generate composite ops with wrapped numbers. For composite ops, we should only need to generate the base ops that compose them.However, in the case of .Scalar ops where the scalar will be promoted to a tensor, we have an issue. The promoted tensor is not on the lazy device by default.
Here, we add guards to first determine if a given tensor is a promoted wrapped number. If so, we create a LazyTensor out of it. Otherwise, we fallback to the normal path.
Test Plan:
lazy_tensor_core/test/cpp/build/test_ptltc --gtest_filter=AtenLtcTsTensorTest.TestAdd*;AtenLtcTsTensorTest.TestBitwiseAnd*