fix: isolate Capture codec fixture in adapter tests

[abrichr]

Summary

• replace the Capture adapter tests' machine-global codec dependency with a
  deterministic, lossless test-only writer/extractor boundary
• continue exercising Capture's real SQLAlchemy schema, CaptureSession
  action processing and public frame lookup, the Flow conversion bridge, and
  the unmodified compiler
• keep codec integration qualification in openadapt-capture, where the
  external FFmpeg contract is owned

Root cause

Flow v1.20.1 qualification resolved openadapt-capture==1.0.2. After
openadapt-capture==1.0.3 published, release-commit CI re-resolved the floating
capture extra and all Capture adapter cases failed closed with
FFmpegUnavailableError. Capture 1.0.4 also changes the implementation to
direct subprocess.Popen streaming with MP4 timing metadata, so emulating
private FFmpeg commands would be brittle.

This fix patches only this test module's VideoWriter name and Capture's public
openadapt_capture.video.extract_frame seam. It does not patch _run_checked,
subprocess.Popen, concat/PNG commands, MP4 UUID internals, or any production
path.

Safety / production impact

• no Capture version pin or upper bound
• no bundled codec binary
• no production-code change
• missing/misconfigured FFmpeg still fails closed in production

Verification

• ruff check tests/test_capture_adapter.py
• ruff format --check tests/test_capture_adapter.py
• pytest -q --timeout=300 tests/test_capture_adapter.py — 34 passed against
  exact Capture main c11969b7efd2eaaf9981d853fbe62d2bdaddd953
  (direct Popen/UUID implementation)
• the same focused suite — 34 passed against the published
  openadapt-capture==1.0.4 wheel
• the same focused suite — 34 passed against the published
  openadapt-capture==1.0.3 wheel
• full GitHub Actions matrix pending on this draft