fix: count queued HLS transcodes

[tcconnally]

Corrective follow-up

This PR corrects the transcode-prefetch accounting from #57.

The prior implementation counted active HLS streams but excluded queued HLS playlist entries. Because mpv can open a queued prefetch demuxer before playback, multiple pending HLS transcodes could still exceed MAX_CONCURRENT_TRANSCODES.

This change:

• counts current and queued HLS loads together;
• applies that count before choosing a transcoded prefetch URL;
• preserves the target cell's queued state during filter changes until mpv replaces the playlist;
• adds regressions proving pending HLS loads consume the budget and that the wall accounting includes them.

Verification

• Reliability tests: 54 passed.
• macOS playback contract tests: 3 passed.
• Playlist tests: 10 passed.
• Full tests/run_all.py: all suites green.
• compileall and git diff --check: passed.
• Independent exact-tree review: passed; prior blocker reproduced and covered by the new tests.

Native macOS/Emby soak remains the target-host acceptance test after pulling the merged result.