Remote upload watch mode was effectively unusable on large workspaces

[voarsh2]

(e.g. ~6K files) on Windows due to two issues:

• The standalone client’s SimpleHashCache reloaded file_cache.json and re-validated every cached path on each access, causing O(N^2) Path.exists() calls and constant file_cache.json rewrites.
• Both standalone and non-standalone clients re-hashed the contents of every candidate file on every watch iteration, even when mtime/size were unchanged.

This change:

• Updates standalone_upload_client.SimpleHashCache to:

  • Load file_cache.json once into memory and run the stale-check only once per process.
  • Keep an in-memory dict for all subsequent get/set/all_paths calls.
  • Stop flushing the cache file on every set/remove, eliminating the continuous rewrites during watch.

• Adds an in-memory stat cache to RemoteUploadClient.detect_file_changes in both scripts/standalone_upload_client.py and scripts/remote_upload_client.py:

  • Tracks (mtime_ns, size) per absolute path.
  • Uses stat() to short-circuit unchanged files and only re-hash files whose stat changed or are new.
  • Keeps initial behavior of treating the first watch iteration as a full sync (all files "created"), but makes subsequent iterations fast and responsive even on large Windows repos.

Result: both standalone and non-standalone --watch modes now detect single-file edits quickly on a ~6K-file Windows workspace without hammering cache files or re-hashing every file on each interval.