[ONNX] Support dynamic scale & zero_point for fake_quantize_per_tensor_affine
Dynamic scale & zero_point requires opset 13 `ONNX::QuantizeLinear`
and `ONNX::DequantizeLinear`.
Improved error message when scale is not constant for opset 10 symbolic function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75697
Approved by: https://github.com/garymm
