Record when each stretch of speech starts and ends, and when a question finishes

The package's onInferenceStart fires when the voice detector closes a stretch of speech; the end is that moment less the detector's 1.4 s wait, and the start is the end less speechMs. Each exchange keeps when it was asked and when the interviewer finished asking it (moved by a repeat the candidate asked for, not by a nudge), for the report's time to first word and pauses.