Fix empty indices for gather op

[BruceXinXin]

Fix empty indices for gather op (torch.embedding zero-dimension crash)

Problem

torch.embedding(weight, indices) should support indices tensors of any shape, including those with a zero-sized dimension. Given:

• weight.shape = [vocab_size, embedding_dim]
• indices.shape = [d1, d2, ..., dn]

The expected output shape is [d1, d2, ..., dn, embedding_dim], and this should hold even when any di = 0 (producing an empty output tensor).

Previously, the lowering of torch.embedding to the gather op would fail or produce incorrect results when any dimension of indices was zero. This is a valid and common case in practice — for example, when batching sequences of variable length that are filtered down to an empty set.

Fix

Handle the empty-indices edge case in the gather op lowering so that when indices contains a zero-sized dimension, the op correctly produces an output tensor of the expected shape with zero elements, rather than crashing or emitting incorrect IR.

Testing

Added a test case covering torch.embedding with zero-sized indices to verify correct shape inference and lowering behavior.

[BruceXinXin]

@sjarus @Lallapallooza

[sjarus]

I've added @tanyokwok - the author of the code - as reviewer. I'm not personally familiar with the StableHLO pathway code.

[BruceXinXin]

Seems @tanyokwok is no longer active:) Do you know any one I can ask for review? Thanks