OptionalgroupVideos only: per-temporal-group timestamps in seconds (frame-pair
averaged, transformers _calculate_timestamps) — the <X.X seconds>
headers the prompt builder renders before each frame group.
Language-side image token count = t * (h/merge) * (w/merge).
[gridTgridHgridW, Ctpsps*ps] f32 patch rows (merge-block-major).
[t, h, w] in PATCH units (h = resizedH/16 etc.).