Add regression test for length-1 tensor int casts

[hayeah]

Summary

• add a Torch frontend regression test for int() on a foldable length-1 tensor
• cover the minimal shape pattern that previously failed in coremltools 9.0
• keep the test conversion-only so it exercises the frontend bug without requiring CoreML runtime prediction

Context

The released 9.0 converter fails on traced Torch graphs that contain aten::Int over a foldable length-1 tensor, for example a model that reshapes with int(h * w) after unpacking x.shape. Current main already handles this correctly in _cast; this test locks that behavior in.

Minimal repro shape pattern:

b, c, h, w = x.shape
return x.reshape(b, c, int(h * w))

I verified locally that:

• coremltools 9.0 fails on this pattern with TypeError: only 0-dimensional arrays can be converted to Python scalars
• current main converts it successfully

[TobyRoseman]

Your unit test passes for me with the 9.0 release.