Fix audio inference - Pass --level 0 to audio feature extractor, and improve error handling

[orangekittysoftware]

I found that the audio inference stage was failing due to a misleading error about Python type mismatches on the 'video' parameter - turned out the .npy file was not being created because the --level 1 default was being interpreted very lierally by the aduo feature extractore script. Passing an explict --level 0 resolved the issue.

Also added some paranoid error checking around the intermediate file creation during audio inference.