fix(preprocess): eliminate MOG2 cold-start frame-0 whiteout

[Chouffe]

Problem

A fresh MOG2 subtractor runs per video, and it has no background model on
its first apply() — so frame 0 of every clip is flagged ~100% motion
and the red (motion) channel whites out, versus a steady-state motion
fraction of ~0.02–0.05.

Fix

Feed frame 0 to MOG2 so it still learns from it, but emit frame 1's
processed output in its place (duplicated), so the whiteout never reaches
the video while the frame count and index alignment stay intact. Applies to
both stabilize_and_preprocess_sonar_video.py and preprocess_sonar_video.py.

Why not "warm up MOG2 with the first N frames": re-emitting the warmup
frames double-counts them and corrupts the background model, inflating
steady-state foreground 30–100%. Priming without re-emitting is clean but
drops N frames. Duplicating frame 1 keeps every other frame untouched.

Verification

stabilize_and_preprocess CLI on small.mp4 (300 frames):

	frames	frame-0 red	steady 10–50
before	302	1.0000 (whiteout)	0.0200
after	302	0.0080	0.0202

Frame count preserved; every frame ≥1 is byte-identical by construction
(MOG2 is fed the same 0,1,2… sequence). ruff clean, 99 tests pass.

Refs: lumax-eco/sonar-smolt-detection#107