Related to #595:

• the model in tweetynet uses the old metrics
• switching to torchmetrics (in ENH: Add ability to compute metrics by accumulating across samples when desired #595) will require releasing a new version of tweetynet with that instead uses the metrics classes torchmetrics
• taken to logical extreme we will end up updating a bunch of packages with every change to vak
• also, it's confusing for a user to need to install an extra package just to have a model

We should instead just essentially cut and paste TweetyNet into vak.models and then use the new metrics.

One question is whether we want a separate nets module/sub-package. This would help with readability of the codebase because there'd be less code per module -- you'd import TweetyNet (the neural network architecture) into vak.models.TweetyNetModel but then the API starts to get verbose from the POV of the outside, and you increase potential for issues like circular imports. I think I'd rather just have a vak.models sub-package only.