Fix squeeze value inference for scalar outputs

[ZedingZhang]

Summary

• normalize scalar results from squeeze.value_inference() to the MIL input dtype
• preserve NumPy scalar values when upstream symbolic value inference produces an object array
• add a regression test for the concat -> split -> squeeze path from the issue

Fixes #1811.

Testing

python -m pytest \
  coremltools/converters/mil/mil/ops/tests/iOS14/test_tensor_transformation.py::TestSqueeze::test_builder_eval_rank_0 \
  coremltools/converters/mil/mil/ops/tests/iOS14/test_tensor_transformation.py::TestSqueeze::test_builder_eval_rank_0_from_split \
  coremltools/converters/mil/mil/ops/tests/iOS14/test_tensor_transformation.py::TestSqueeze::test_squeeze_value_inference_is_inplace -q

Result: 3 passed.

[TobyRoseman]

The unit test should more closely follow the repo code in the GitHub issue. Specifically, it should call ct.convert.

[ZedingZhang]

Updated the regression test to call ct.convert, as requested. This also exposed an object-dtype issue in split value inference during constant elimination, so fully known split outputs are now cast back to the input dtype.

[TobyRoseman]

Also like the repo code, you need to call .predict on the model return by ct.convert.

[ZedingZhang]

Updated. The regression test now calls .predict() on the model returned by ct.convert() with a concrete input and verifies that the scalar output is 1.0 on macOS 12+.