[Feature][Relax][ONNX] Support QLinearMatMul in Relax ONNX frontend #20958

Description

The ONNX QLinearMatMul operator (opsets 10 and 21) is currently missing from the Relax ONNX frontend (python/tvm/relax/frontend/onnx/onnx_frontend.py), as noted in RFC tracking issue #20124 under quantized operators.

QLinearMatMul computes 8-bit quantized matrix multiplication with 32-bit accumulation and dequantization/requantization:

Y = clip(round((A - a_zp) * (B - b_zp) * (s_a * s_b / s_y)) + y_zp, min_val, max_val)

Key Requirements

• Support opset 10 (uint8) and opset 21 (int8, uint8, int16, uint16).
• 32-bit integer accumulation to prevent intermediate arithmetic overflow.
• Per-row and per-column zero-point and scale broadcasting.
• Correct rounding and clipping to the output data type range.
• Comprehensive unit tests verified against ONNX Runtime.

Reference PR

Implementation is available in PR #20951.

Triage

• needs-triage
• frontend:onnx