[kitlangton]

On Windows, a model download now waits for one another process is already running instead of failing at once.
This also fixes waiting_for_another_processs_download_reports_its_partial_bytes, which has failed on every Windows check run since #1735 added it.

❌ Before

1. hex model install is downloading a model in a terminal (or another Hex is).
2. Hex starts downloading too, and finds the download lock already held.
3. ❌ On Windows the lock call returns ERROR_LOCK_VIOLATION, which Rust's io doesn't map to WouldBlock, so Hex took it as a real failure: the download failed rather than waiting, and deleting a model during a download failed with a raw OS error instead of "A model download is in progress".

✅ After

1. Same.
2. Same.
3. ✅ Both lock calls use the check meetings already used (WouldBlock or ERROR_LOCK_VIOLATION), so Windows waits and shows the other process's progress, as macOS and Linux did.

- Err(e) if e.kind() == WouldBlock => wait     // never true on Windows
+ Err(e) if lock_contended(&e) => wait         // one check, moved to durable.rs

👀 Read: nothing tricky; meeting.rs's lock_contended moved to durable.rs unchanged so models and meetings share it.

Checks run

• scripts/check transcription::models meeting (macOS): 383 passed
• check-windows.yml on this branch: passed on x64 and ARM64, https://github.com/anomalyco/hex/actions/runs/43303190525 (ARM64 needed one rerun: an unrelated Authenticode test timed out)