Losslessly trim every track to a time range [start_sec, end_sec).

Per track: select samples by decode time (snap the start back to the nearest keyframe), rebuild all sample tables from the kept set, and re-base the timeline to zero. Kept chunk-runs from all tracks are written to a new mdat sorted by their original file offset, preserving A/V interleave. Duration headers are updated. Memory-safe: the new mdat is a segment list.