Use the first dim of size 3 when torch.cross has no dim

[LeSingh1]

dim is optional in aten::cross. Omitted, torch uses the first dimension of size 3, which is not dim 0 in general. The converter passed the argument straight to mb.gather, where the missing dim fell back to axis 0.

class M(nn.Module):
    def forward(self, x, y):
        return torch.cross(x, y)          # no dim

a = b = torch.rand(2, 3)

torch returns shape (2, 3); Core ML returns (3, 3) — wrong shape and wrong values, with nothing raised. Passing dim=1 explicitly has always been correct, which is why the existing test_cross (which always passes dim) never caught it.

The fix resolves the default the way torch does, and raises when no dim has size 3 rather than gathering along a dim that cannot hold a cross product.

Testing

TestCross::test_cross_default_dim over shapes (2, 3), (3, 4), (3, 3), (2, 3, 4). The (2, 3) and (2, 3, 4) cases fail on main; (3, 4) and (3, 3) pass either way and are there as controls, since for those the first dim of size 3 is dim 0. Existing test_cross is unchanged and still passes.

This machine has a broken scikit-learn install that makes every TorchFrontend.TORCHEXPORT case error out, including pre-existing ones, so I could only exercise the TorchScript frontend locally.