[serialize] Support torch.gather export via Circle GatherNd

[mhs4670go]

This commit supports torch.gather export via Circle GatherNd.

TICO-DCO-1.0-Signed-off-by: seongwoo mhs4670go@naver.com

[mhs4670go]

@arkq

I agree with the overall lowering direction in #786: general torch.gather is not a good fit for Circle GATHER, and constructing full-coordinate indices for Circle GATHER_ND matches the expected semantics.

However, I am concerned about the PR's implementation strategy of keeping the original aten.gather node and changing its index input to a full-coordinate tensor, then relying on a metadata flag to make the serializer interpret it as GATHER_ND.

After this pass, the node no longer has valid aten.gather semantics. In other words, the graph still says aten.gather(input, dim, index), but index is now shaped like [..., rank] and is meant to be consumed by GATHER_ND. This makes the intermediate IR misleading and invalid from a PyTorch/ATen point of view. It also means later passes, meta propagation, debugging, or future optimizations may accidentally treat this node as a normal aten.gather, even though its actual meaning has already changed.

I think it would be cleaner and safer to materialize an explicit internal GatherNd-like representation instead:

aten.gather(input, dim, index)
  -> build full-coordinate indices
  -> internal_gather_nd(input, full_indices)
  -> serialize internal_gather_nd as Circle GATHER_ND

This would make the semantic boundary clear:

• before the pass: the graph has PyTorch gather semantics
• after the pass: the graph has explicit Circle/TFLite-style GATHER_ND semantics
• the serializer no longer needs to infer a semantic change from a metadata flag

Because of this, I would prefer not to merge #786 in its current form. Instead, I suggest we keep the same high-level idea—lowering torch.gather to Circle GATHER_ND, which is why I posted this PR.

Could you review this PR instead?

[arkq]

I am concerned about the PR's implementation strategy of keeping the original aten.gather node and changing its index input to a full-coordinate tensor, then relying on a metadata flag to make the serializer interpret it as GATHER_ND.
After this pass, the node no longer has valid aten.gather semantics.

Yes, I totally agree with you. I did not know how to implement torch.gather to circle.gather_nd conversion while preserving torch graph validity - I was not aware that there is torch.ops.circle_custom.gather_nd. Your approach seems to be better indeed. The code looks OK.