Q/DQ and QONNX parsing, GPU INT8 and FP8 GEMM/MatMul

[shaunlee8]

First-class quantization and low-precision pipeline for SOFIE. Covers GEMM/MatMul operators from Q/DQ ONNX and QONNX graphs. Detects GEMM/MatMul regions by inspecting local Q nodes after the base graph is built.

• Support for QONNX parsing and Q/DQ ONNX parsing.
• Simple CPU lowering path for quantized GEMM added.
• Per-tensor and per-channel weight quantization support.
• Generate optimized INT8 and FP8 GEMM and MatMul calls through cuBLASLt.
• cuBLASLt exact, padded, flattened tensor shapes supported.
• External binary weight storage for INT8/FP8.
• Test coverage for QONNX, Q/DQ, codegen, and GEMM/MatMul execution.