Force decode timestamp after speaker turn

[akashmjn]

Fix for #10 by updating the logit filter ApplyTimestampRules in decoding.py.

Re-ran the run_pipelines.py and updated the analysis notebook & metrics. Numbers for tdrz didn't really change much.

Also includes the following minor improvements:

• Fix a small bug in run_pipelines.py
• Decoding of <|speakerturn|> tokens is also now also configurable via an added DecodingOption e.g. whisper.transcribe(with_speaker_turns=False).
• Speaker turns are better formatted in verbose output, and also included in written JSON segments in field before_speaker_turn