Adaptive loop-filter reconstructed model is incorrect

For the adaptive LOP3 tool, the model reconstructed at the decoder end differs than the one used at the encoder end (checksums). The decoder crashes when the model is run.

The adaptive loop filters use a custom conversion script from onnx to sadl that extracts the base model quantisers and runs the model quantisation from float to int16. This script is behind the one in sadl v10 and might be the reason for the issue.

This affects NNVC v10rc