Separate severity from confidence level

[finnbear]

The filter should be be able to output/filter severity independent of confidence. It should be possible to represent a low-confidence, severe profanity detection.

Would help solve #34