Fix incorrect word timestamps for audio inputs longer than 60 seconds

[JamesVong]

Bug

ExtractWordsFromAudio produces incorrect word timestamps for any input longer than 60 seconds (the ChunkEvents max_duration threshold).

ChunkEvents splits long audio into chunks that reuse the same WAV file with a non-zero offset, e.g. for a 96s video:

	start	offset	filepath
Chunk 1	0	0	audio.wav
Chunk 2	60	60	audio.wav

WhisperX is called once on the full file, so word times are already relative to the file start. The existing formula:

transcript["start"] += audio_event.start + audio_event.offset 

double-counts the offset for chunk 2 (adds 120s instead of 0), shifting words at 90s → 210s.

This also inflates the CategoricalEvent duration computed in get_loaders, causing list_segments to generate ~2× too many segments, so a 96s video produces 215 segments spanning 0–214s instead of ~96.

Fix

transcript["start"] += audio_event.start - audio_event.offset

This correctly converts from file-relative time to timeline-absolute time. For chunk 2 (start=60, offset=60): word at 90s → 90 + 60 − 60 = 90s ✓

Chunk 1 (start=0, offset=0) is unaffected.

Reproduction

I ran the events pipeline on a 96s video to confirm the fix:

Chunk (start=0.0, offset=0.0):
  OLD: word times 0.071s – 93.702s
  NEW: word times 0.071s – 93.702s   ← unchanged

Chunk (start=60.0, offset=60.0):
  OLD: word times 120.071s – 213.702s  ← wrong
  NEW: word times 0.071s –  93.702s   ← correct

The demo notebook (tribe_demo.ipynb) uses a 52s Sintel clip which never triggers chunking, so this bug is not visible in the default demo.