fix: optimize serve/sourcemap startup from >180s to ~8s

Root cause: notify-debouncer-full with macos_kqueue caused 25K+
directory watch registrations to take >170s on large projects.

Changes:
- Replace notify-debouncer-full with notify-debouncer-mini to avoid
  FSEvents Create+Remove event cancellation and heavy per-watch overhead
- Remove macos_kqueue feature; use FSEvents (matches upstream Rojo)
- Fix VFS read guard: watch_enabled && watch_recursive -> watch_enabled
  so recording mechanism works when watch_recursive is false
- Build RefPathIndex from patch_apply data instead of redundant 4.8s
  WalkDir + 5,854 file re-reads (populate_from_dir)
- Use ServeSession::new_oneshot() for sourcemap without --watch
- Restore simple synchronous watch setup (implicit during reads)

Serve: >180s (never completed) -> ~8s to Listening
Sourcemap: 14.3s -> ~9s
All 489 tests passing.

crates/memofs/Cargo.toml

@@ -18,14 +18,9 @@ homepage = "https://github.com/UserGeneratedLLC/rojo/tree/master/memofs"
 crossbeam-channel = "0.5.15"
 fs-err = "3"
 log = "0.4"
-notify-debouncer-full = "0.7"
-serde = { version = "1.0.228", features = ["derive"] }
-
-[target.'cfg(target_os = "macos")'.dependencies]
-notify = { version = "8.2", features = ["macos_kqueue"] }
-
-[target.'cfg(not(target_os = "macos"))'.dependencies]
 notify = "8.2"
+notify-debouncer-mini = "0.7"
+serde = { version = "1.0.228", features = ["derive"] }
 
 [dev-dependencies]
 tempfile = "3.26.0"