bug: v0.2.1 — upgrade cleanup, online-file races, instant exit

[Gumbees]

Three upgrade/lifecycle bugs in one patch. Cargo bumped to 0.2.1.

1. Manual upgrade from 0.1.x left leftovers + could deadlock migration

The in-app updater rename-dances on the installed binary, but a user who downloads the new release manually from GitHub Releases (or upgrades across the 0.2.0 install-layout change) ran into:

• Old binary still on disk at %APPDATA%\ImmichSync\ or %APPDATA%\bees-roadhouse\immichsync\.
• Stale HKCU\...\Run\ImmichSync autostart entry pointing at the old path → both 0.1.x and 0.2.x launch at next login.
• Stale Apps & Features Uninstall key still pointing at the old install dir.
• Worst: if v0.1.x was still running, migrate_to_split_layout tried to rename an open state.db → silently fell back to copy → remove_file failed → divergent DBs at old + new paths.

Adds cleanup_legacy_install() (called from main() before the migrations) that:

• Detects legacy binaries at both known 0.1.x paths.
• taskkill /F /FI 'PID ne <ours>' /IM immichsync.exe to release file locks.
• Removes the legacy binary + version.txt.
• Clears the Run / Uninstall registry entries only when their value points at a legacy path (users with autostart disabled stay disabled).

2. SHA-1 hash + upload-stream hydrated cloud placeholders

The watcher's filter drops cloud-placeholder files at discovery, but a file can be evicted between watcher → queue or queue → worker (OneDrive / SeaDrive auto-evict under storage pressure). Reading the placeholder triggered a full cloud download — gigabytes of bandwidth + CPU for a file we were about to skip anyway.

The comment at queue.rs:233-237 explicitly relied on the watcher's filter being authoritative; it's not, race-wise.

• queue.rs:process_file: is_online_file(&meta) re-check between fs::metadata and hash_file. Skips with Ok(None).
• worker.rs:process_item: is_online_file re-check after pickup, before opening the file for upload. Marks the entry "failed" with a recognizable message; the watcher's reconciliation re-enqueues when the file is local again.

3. Tray Quit / WM_QUIT waited up to 3s

pipeline.stop() graceful path awaits the spawned per-item upload tasks (up to 3s). On user-initiated Quit that's a noticeable hang — "forget the current operation, drop it, exit."

Adds App::shutdown_and_exit() -> !:

• Stops the watch engine (releases file handles, fast).
• pipeline.signal_stop() — signals the worker but doesn't wait.
• std::process::exit(0).

In-flight `"uploading"` rows are recovered by `reset_stale_uploading` on next launch (existing mechanism), so partial uploads retry cleanly. The graceful shutdown() path is preserved for the RestartToUpdate flow (the new instance needs a clean handoff of state.db).

Test plan

• `cargo test` — 110 tests passing locally (added 5 new)
• `cargo clippy --all-targets -- -D warnings` clean
• `cargo fmt --all -- --check` clean
• Manual: install v0.1.8, then launch a v0.2.1 binary from Downloads. Verify legacy binary + Run entry + Uninstall key are gone after first launch.
• Manual: SeaDrive evict a tracked file mid-queue, verify worker skips it without hydrating.
• Manual: click tray Quit while an upload is in-flight, verify tray disappears instantly.