Add 2-bit quantization support to WebGPU GatherBlockQuantized operator

[Shivani767]

Adds complete 2‑bit quantization support to the WebGPU GatherBlockQuantized operator. The implementation includes:

1. Sign‑extension logic for 2‑bit signed quantized values
2. Sign‑extension logic for 2‑bit signed zero points
3. Maintains full backward compatibility with existing 4‑bit and 8‑bit support

Motivation and Context

This change addresses GitHub issue #28895: "Add 2‑bit quantization support to the WebGPU GatherBlockQuantized operator". The existing codebase had partial 2‑bit support but was missing sign‑extension, which is critical for correct dequantization of signed 2‑bit values (range [-2, 1]).

[Shivani767]

Thank you for the review and merge.

This contribution adds complete 2-bit quantization support to the WebGPU GatherBlockQuantized operator by implementing proper sign-extension for signed 2-bit values and zero points, ensuring correct dequantization behavior while maintaining compatibility with existing 4-bit and 8-bit paths.

Appreciate the feedback and support from the ONNX Runtime maintainers.