feat(diags): time the host render path in video_render_probe

[kfox]

Why

video_render_probe.py modelled what a frame costs to put on the wire but measured nothing about what it costs to produce. That leaves it unable to answer whether a given machine is fast enough to drive c64cast at all — the question that decides whether the host can be a small single-board computer instead of a laptop.

What

• Times decode and render separately and reports mean / median / p95 / max ms per frame alongside the existing per-region write cost.
• A verdict line naming which side, if either, actually binds the source frame rate. It compares both ceilings against the source rate, so a clip both sides clear comfortably reports "neither side binds" rather than misreading headroom as a bottleneck.
• --threads N pins decode and OpenCV to N threads, so two machines can be compared by single-core speed rather than by core count.
• Two new per-frame CSV columns: decode_ms, render_ms.

The frame loop now steps the decode generator by hand instead of using a for ... in container.decode(v), so the generator advance — where the decoder actually runs — can be timed apart from the render path it feeds.

The measurement this makes visible

Compose cost tracks the source resolution, not the display mode, because every mode calls cv2.resize() down to its own small target and that resize reads every source pixel:

Source	mhires	hires	mcm	petscii
3840x2160	34.3 ms	33.4	30.4	29.8
1920x1080	11.2 ms	10.4	8.0	7.6
1280x720	6.7 ms	6.0	3.8	3.4
320x200	3.0 ms	2.1	0.4	0.1

So the media, not the renderer, is usually what decides whether the host or the link is the bottleneck — which is why the verdict names pre-scaling as the fix rather than just printing a ratio.

Verification

• Run across three source resolutions (4K, 1080p, 320x200) confirming both verdict branches and the --threads pin.
• make check (ruff, pyright, 3801 tests) and make site-check pass.
• Offline only — no hardware involved.

Two pre-existing ruff format findings in docs/extending.md are untouched by this branch (they fail identically with these changes stashed).