Convert float exponents in aten.pow to int when possible

[zjgarvey]

Pure floating point pow operations no-longer support negative base values (see llvm/llvm-project#126338), but many models coming from ONNX use floating point representations of integers as the exponent.

This change:

1. matches on constant rank-0 exponents and converts them to scalar constants.
2. matches on constant floating-point scalar exponents and converts them to ints if possible.
3. lowers Tensor(float)^int cases to math.fpowi

Addresses some remaining test failures related to iree-org/iree#19996.

[praveen-g-ctt]

LGTM