_modernize_ timeframe-labeling evaluation with `common` package

[keighrim]

Currently, the evaluation code depends on pyannote's segmentation-based metrics as accuracy measures. That said, I want to investigate the following possibilities;

1. timeframe-labeling can be merged with forcedalignment evaluation.
2. namedentity-recognotion is also a sequential tagging problems, and hence can be merged in FA and TFLabeling evaluation.

[shel-ho]

Upon investigation,

1. TimeFrameLabeling (TFL) cannot be merged with ForcedAlignment (FA) for the following reasons:
   1. Differing metrics: FA uses segmentation coverage/purity while TFL uses DER and detection precision, recall, and F1.
   2. Differing gold csv formats: FA has a '\t' delimiter while TFL has a ',' delimiter.
   3. Differing inputs: TFL takes video documents and will throw an error if there are no video documents while FA only takes text and audio documents.
   4. FA requires conversion from cadettimes to ms while TFL is already in ms.
2. NamedEntityRecognition (NER) cannot be merged into TFL for the following reasons:
   1. NER labels tokens and not time frames
   2. The conventional method of evaluating NER where every partial match is both a false positive and false negative is not compatible with pyannote metrics.

[keighrim]

During meeting today, we discussed the possibility to pick one metrics for all TF labeling task, regardless of the labeling is binary or multi-class. @shel-ho to work on researching the available classes in pyannote libray.