Stop ImageWatcher per-file processing from blocking the event loop

[thusser]

Implements specs/plans/2026-08-20-imagewatcher-event-loop-blocking.md (Status stays proposed — the plan's own checklist flips it to implemented only once landed and verified at the site over one night).

Problem

ImageWatcher._worker processes every watched file with synchronous, never-yielding operations directly on the module's single event loop:

1. fits.HDUList.fromstring(data) — a synchronous, CPU-bound astropy parse per file.
2. LocalFile.read / write / close / remove / exists / find — declared async def but execute plain blocking syscalls with no executor (listdir already offloads; the others never got the treatment).
3. LocalFile.__init__ opens the file synchronously, and VirtualFileSystem.open_file has no await point before that — so even with read/write/remove offloaded, a slow/network-mounted watch path could still stall the loop at open time.

Result: RPC replies, inotify handling, and comm keepalive freeze for seconds at a time (the 2026-08-20 MONET South incident, module.py:205/207 "Event loop stalled for 0.81s" / 2.81 s total).

Changes

• pyobs/modules/image/imagewatcher.py: offload the FITS parse via asyncio.to_thread in _worker; add docstring notes that process_extra/cleanup_extra run on the event loop and must not block.
• pyobs/vfs/localfile.py: route read, write, close, remove, find, exists through the default executor (module-level _find_sync/_remove_sync helpers; exists calls os.path.exists directly), mirroring the existing listdir pattern. Move the open()/makedirs work out of __init__ into an overridden __aenter__ via a _open_sync helper — path validation still raises synchronously at construction; every call site already uses async with, so nothing observes the open moving one step later. All VFS consumers inherit the fix.
• Tests: heartbeat-style loop-responsiveness tests in tests/modules/image/test_imagewatcher.py (slow FITS parse) and tests/vfs/test_localfile.py (slow read/write/open via a fake slow fd and a monkeypatched _open_sync). Verified they fail on the old code and pass on the new.

Verification

• pytest tests/modules/image/ tests/vfs/ → 34 passed, 1 skipped (pre-existing SSH-disabled skip).
• Full suite: 1558 passed; the 14 failures are sandbox-environment issues unrelated to this change (read-only /opt/pyobs/storage, read-only ~/.config/sunpy, missing pyobs console script in the venv).
• ruff clean, black clean, pyrefly 0 errors on touched files.

Out of scope (per plan)

Archive upload latency / queue backlog (already async, never freezes the loop) and step 0's live site measurement are flagged for follow-up in the plan.