onnx: import GatherBlockQuantized without widening the table

[czoli1976]

Adds the com.microsoft.GatherBlockQuantized importer. ORT-GenAI exports look up their embedding through it, so without a handler those models cannot be imported at all.

Approach

The rows are gathered before being dequantized, so the table stays 4-bit in memory and only the selected rows are widened. Dequantizing the table up front — the straightforward lowering — would cost eight times its size for a vocabulary-sized embedding: on a 1.7B Qwen3 export that is 155 MB of int4 turning into 1.24 GB of f32, for a step that only ever reads a handful of rows.

Everything is expressed with existing ops, so there is no new eval kernel: Gather on the packed table, a nibble split (the values are integral and below 256, so /16 and the remainder are exact in f32), a per-block broadcast of the scales, and (value - zero) * scale.

Scope

Rank-2 table, bits=4, uint8 storage, gather_axis 0 (which the operator requires for uint8 anyway) and the last axis as quantize_axis — what the exports emit. Anything else is rejected with a specific message rather than silently mis-lowered. Absent zero_points, the zero defaults to 8 as the spec states.

Testing

Checked against onnxruntime 1.28 across block sizes 16 / 32 / 128, with and without zero points, and for 1-D and 2-D index shapes: bit-exact in every case (the arithmetic is exact, so this is equality, not a tolerance). A case is added under onnx/test_cases/ with onnxruntime-generated expectations, passing all four passes including the NNEF round-trip.

Note that onnxruntime only registers this op in reasonably recent builds — 1.19 is too old to run generate_io.py.

🍍