[ONNX] Add infra for quantized model export and support quantized mobilenet v3 (#72215)
* Add infrastructure and helper functions to enable future work for other quantized operators and models.
* Add export for quantized operators needed by torchvision mobilenet v3 large.
* ATen namespace: hardsigmoid, flatten, adaptive_avg_pool, quantize_per_tensor, dequantize.
* Quantized namespace: conv2d, conv2d_relu, hardswish, add, mul.
* Numerous bug fixes, in unpack_quantized_weight.cpp, symbolic functions, and unit test.
Co-authored-by: BowenBao <bowbaomicrosoft.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73102
