addon.node: VAD-aligned (faster-whisper-like) timeline with real gaps

[buxuku]

Summary

With VAD enabled, whisper.cpp concatenates all detected speech into a single stream, so addon.node returned a gap-less timeline where every segment end equals the next segment start. This PR adds an alignment layer that puts timestamps back on the original timeline with real silence gaps, similar to faster-whisper.

New params (all optional, backward compatible)

param	default	meaning
align_mode	"hybrid"	hybrid | run | word | legacy
vad_merge_gap_ms	2000	adjacent VAD segments whose silence gap is ≤ this (ms) merge into one decode "run"; a larger gap becomes a real subtitle gap. Negative disables aligned mode.
word_gap_ms	500	(word/hybrid) start a new segment when the gap between two consecutive words exceeds this (ms).

Modes

mode	gaps between regions	word-level sub-split	boundary accuracy	notes
hybrid (default)	yes	yes	best	VAD-grouped runs, each run sliced into its own buffer and re-segmented by word gaps; every segment end clamped to its last word
run	yes	no	good	per-run decode emitting whisper's own segments
word	yes	yes	ok	single decode pass + word-gap re-segmentation; uses core VAD via the new whisper_full_get_token_t0/t1 mapping when a VAD model is given
legacy	no (original)	no	–	original continuous single pass

Implementation notes

• Each run is decoded from a physically sliced audio buffer, not via offset_ms/duration_ms (those only bound the outer seek loop and let neighbouring speech bleed into short runs). Slice-relative timestamps are shifted back onto the original timeline with a per-run base offset.
• Progress is rescaled across runs so the JS progress callback still sees a single monotonic 0..100.
• VAD context is forced onto CPU (matches whisper.cpp internals; the Metal VAD graph aborts).
• Depends on the core getters whisper_full_get_token_t0/t1 added in this branch.

Test plan

Verified on 4 synthetic audios (multi-gap / sub-split / leading+trailing silence / choppy) × 4 modes:

• big gaps (> vad_merge_gap_ms) become real empty spans;
• no text is emitted inside silence (incl. leading/trailing);
• hybrid has the most accurate boundaries;
• legacy stays continuous (unchanged behavior).