Fix symbolic nearest1d upsample output rank

[ZedingZhang]

Summary

• preserve statically known output dimensions in the Torch upsample dialect shape inference
• keep genuinely dynamic output dimensions symbolic
• add a TorchExport regression test for symbolic nearest-neighbor 1D interpolation followed by Conv1d
• add direct shape-inference coverage for both nearest and bilinear Torch dialect ops

Root cause

The PyTorch frontend implements 1D nearest-neighbor interpolation by expanding the input with a dummy width dimension, invoking the 2D Torch upsample dialect op with output_width=1, and squeezing that dimension afterward.

The dialect op previously replaced both output dimensions with fresh symbols unconditionally. As a result, the dummy width was inferred as symbolic even though it was the constant 1. The following squeeze could not remove it during type inference, so Conv1d received a rank-4 input and failed conversion.

This change preserves the concrete value when an output dimension is known, while continuing to create a symbol for a dynamic dimension.

Tests

python -m pytest \
  coremltools/converters/mil/mil/tests/test_programs.py::TestMILBasic::test_get_dialect_namespaces \
  coremltools/converters/mil/mil/tests/test_programs.py::TestMILBasic::test_torch_upsample_preserves_static_output_dimension \
  coremltools/converters/mil/mil/tests/test_programs.py::TestMILBasic::test_invalid_dialect_namespaces_error_out \
  coremltools/converters/mil/frontend/torch/test/test_torch_ops.py::TestUpsample::test_upsample_nearest1d_with_symbolic_output_size_followed_by_conv1d -q

5 passed

The new end-to-end regression test fails with the reported convolution shape error before the fix and passes afterward. It also asserts that the generated MIL convolution receives a rank-3 input.

Fixes #2837