DOC: Add note/tip about amount of silence when training TweetyNet model

[NickleDave]

Is your feature request related to a problem? Please describe.
Several users now have reported training models will predict only the background label class.
This seems to occur when the training data contains a lot of silent periods.
I have not confirmed this experimentally, but my intuition for what's going on is that it's due to class imbalance, the model basically learns it can perform well by always predicting the background class.

Describe the solution you'd like
There should be at least one place in the documentation where this is mentioned. Places I can think of

• A vignette on training TweetyNet
• A "tips and tricks" section or the FAQ (maybe FAQ should link to "tips and tricks"
• A note/tip in the intro tutorial that links to all of the above