Support a symbolic batch dim in torch narrow and group_norm

[LeSingh1]

Both converters bake x.shape into a constant, so a symbolic dim aborts the conversion with ValueError: Cannot add const [is0, 2, 4] ... cannot have symbolic values.

model = torch.jit.trace(nn.GroupNorm(2, 4).eval(), torch.rand(2, 4, 8, 8))
ct.convert(model, inputs=[ct.TensorType(shape=(ct.RangeDim(1, 8), 4, 8, 8))])

narrow

end = list(x.shape), with only the narrowed dim overwritten. The end of every other dim is irrelevant — the slice covers them whole — so state end = 0 and mask them off instead. That drops the dependency on x.shape entirely. The negative-start handling is unchanged.

_group_norm_impl

The guard was if not any_symbolic(x.shape[2:]), i.e. it checked only the spatial dims, but the static branch also bakes in the batch dim via new_shape = [n, num_groups, c // num_groups]. A symbolic batch therefore took the static path and produced a constant holding a symbol. Widened the guard to the whole shape, and the dynamic path now reads the batch dim from mb.shape rather than from x.shape[0].

The existing test_groupnorm_dynamic makes height and width dynamic but pins the batch to 6, which is why only the spatial dims were covered.

Testing

TestNarrow::test_narrow_dynamic_batch and TestGroupNorm::test_groupnorm_dynamic_batch (ranks 2, 3 and 4, so both the presence and absence of spatial dims are covered). Both fail on main and pass here; existing TestNarrow and TestGroupNorm tests are unchanged and still pass.

I could only exercise the TorchScript frontend locally — this machine has a broken scikit-learn install that makes every TorchFrontend.TORCHEXPORT case error out, including the pre-existing test_groupnorm_dynamic, so those results are not mine to report.