[Feature][Relax][ONNX] Support QLinearConv in Relax ONNX frontend #20959

Description

The ONNX QLinearConv operator (opset 10) is currently missing from the Relax ONNX frontend (python/tvm/relax/frontend/onnx/onnx_frontend.py), as tracked in RFC issue #20124 under quantization operators.

QLinearConv performs 1D, 2D, or 3D quantized convolution on 8-bit inputs and weights, supporting bias and requantizing to the output quantization parameters:

Y = clip(round((Conv(X - x_zp, W - w_zp) + B * s_x * s_w) / s_y) + y_zp, min_val, max_val)

Key Requirements

• Support 1D (conv1d), 2D (conv2d), and 3D (conv3d) convolutions.
• Dequantize inputs and weights with per-channel or scalar scales and zero points.
• Support optional quantized bias B with scale s_x * s_w and zero point 0.
• Support auto_pad modes (SAME_UPPER, SAME_LOWER, VALID, NOTSET), strides, dilations, and groups.
• Numerical correctness verified against ONNX Runtime.

Reference PR

Implementation is available in PR #20953.

Triage

• needs-triage
• frontend:onnx