ENH: add segmental F1 and related metrics

[NickleDave]

from this paper, as implemented in this repo

• add ("macro") segmental F1 metric
• also add per-class segmental precision, recall and F1

[NickleDave]

Note that the linked implementation does not operate on batches, and that it expects numpy arrays, not torch tensors; so a complete translation will use tensors instead, accepting batches and iterating over each item in the batch. I don't think this can be vectorized (very easily) since segments need to be recovered from each item in the batch, which almost always results in ragged / nested tensors