1.3.9-beta release - #19

…reamController#19)

close() (load/teardown threads via close_resources()) nulled
self.video_cache while a render tick could be between its multiple
video_cache reads -> AttributeError on None (logged via the
action-callback catch). Worse than the logged error: a get_frame() that
merely STARTS after the reader is released can resurrect a
cv2.VideoCapture through Mp4FrameCache._maybe_adopt_shared_cache and
leak it, so a bare snapshot isn't enough. Add a per-instance lock with a
check-then-hold pattern: an unlocked None peek keeps post-close
stragglers free, the body holds the lock for the whole pick+decode so
close() waits for an in-flight frame, and no frame begins against a
released reader. Uncontended acquire per tick is noise next to the
frame decode; close() stays idempotent.

Test: tests/scenario_keyvideo_close_race.py (150-round close-vs-tick
hammer over a sleep-widened stub cache: no exceptions, zero cache
accesses after release, quiet None afterwards; plus a real
mp4_tile_cache-registry InputVideo closed under concurrent ticks).
scenario_keyvideo_build's __new__ factory sets the new lock.