Update GatherBlockQuantized to support 2-bits

[HectorSVC]

Description

Update GatherBlockQuantized to support 2-bits.
Updated op schema, implemented the CPU and WebGPU EP.
This helps to make the model smaller.

[Copilot]

Pull request overview

This PR extends GatherBlockQuantized to support 2-bit uint8-packed data, updating schema/docs, CPU/WebGPU implementations, and tests.

Changes:

• Adds 2-bit packing/dequantization support for CPU and WebGPU paths.
• Updates contrib operator schema and generated docs to include 2-bit uint8 support.
• Adds CPU and WebGPU tests for 2-bit uint8 no-zero-point cases.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
onnxruntime/contrib_ops/cpu/quantization/gather_block_quantized.cc	Adds CPU 2-bit uint8 extraction and default zero-point handling.
onnxruntime/contrib_ops/webgpu/quantization/gather_block_quantized.cc	Adds WebGPU shader logic and shape handling for 2-bit packed uint8 data.
onnxruntime/contrib_ops/webgpu/quantization/gather_block_quantized.h	Expands WebGPU bits validation to include 2.
onnxruntime/core/graph/contrib_ops/contrib_defs.cc	Updates schema documentation and zero-point shape inference for packed components.
docs/ContribOperators.md	Updates public operator documentation for 2-bit support.
onnxruntime/test/contrib_ops/gather_block_quantized_op_test.cc	Adds 2-bit packing helper logic and new CPU/WebGPU test cases.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[tianleiwu]

Review Summary

The core zero-point indexing issues from the prior review iteration have been properly fixed in both CPU and WebGPU paths. The 2-bit extraction logic and the scale_row / q_in_row decomposition for zero-point addressing are correct.

Two suggestions:

1. Inconsistent is_int8 vs is_uint8 guards — The 2-bit code paths in ComputeInternal mix is_int8 (which covers both INT8 and UINT8) and is_uint8 inconsistently. Since bits==2 is only valid for uint8, using is_uint8 throughout would make intent clearer.
2. Missing test coverage for 2-bit zero_points — Both new tests only exercise the default-zero-point path. The 2-bit zero-point unpacking logic (the most complex new code) has no dedicated test.