Fix Scheduler freeze when SQL Server has downtime

[jduust]

Summary

Fixes the scheduler freeze that occurs when SQL Server has downtime or restarts (issues #152 and #106).

Root cause

The Tk after-chain in run_tab.loop() died on any unhandled DB exception, leaving the scheduler frozen with no further polls scheduled. SQLAlchemy's pooled connections also went stale across server restarts (no pool_pre_ping).

Changes

• db_util.connect(): enables pool_pre_ping=True and pool_recycle=1800. Engine is kept alive on operational probe failure so it can self-heal via pool_pre_ping on the next session checkout. A 5 s ODBC login timeout is set only for mssql+pyodbc connection strings (driver-specific via a helper); other drivers use their defaults.
• run_tab.loop(): body wrapped in try/except/finally so the app.after reschedule always runs. Exponential backoff 12s -> 24s -> ... -> 600s (cap), counter resets on the next successful tick. Tracebacks appended to a desktop log file so failures that happened overnight are visible afterwards.
• run_tab.check_heartbeats(): per-job DB calls wrapped so one job's failure doesn't abort the whole sweep; first DB error is re-raised so loop() enters the backoff path.
• runner + new scheduler/inflight.py: trigger IDs this scheduler owns are persisted to %APPDATA%/OpenOrchestrator/inflight.json. On Run click, reconcile_orphans() reads the file and marks any still-RUNNING triggers as FAILED with a log entry. Closes the gap where a crash between begin_*_trigger() and the job ending leaves a trigger stuck in RUNNING.

How I tested

• pylint --rcfile=.pylintrc OpenOrchestrator --good-names=OpenOrchestrator: 10.00/10
• flake8 --extend-ignore=E501,E251 OpenOrchestrator: clean
• python -m unittest discover against sqlite: passes
• Manual chaos test on a robot machine: Windows Firewall blocked outbound TCP/1433 to the SQL Server for 15s/60s/90s. Scheduler logged LOST DATABASE CONNECTION, backoff progression, then Database connection restored after N failed attempt(s). No freeze, no manual restart required.

Things worth discussing in review

• The desktop crash log path (~/Desktop/OpenOrchestrator_scheduler_errors.log) is opinionated. Happy to switch to Python's logging module / a configurable path if preferred.
• pool_recycle=1800 is a default that should be fine for most setups; can be made configurable if needed.
• The orphan reconciliation is per-machine via a local JSON file. A future cleaner approach would be a started_by column on the Triggers table so reconciliation can happen DB-side without local state — out of scope for this PR.

[jduust]

Also if you were in doubt, yes this is mostly generated with AI. However we've been running a slightly more barebones "nonlinted" version of the code on one of our machines, and it seems to work as intended so far. Use it if it makes sense, it's a huge gamechanger for us in terms of uptime and catching processes stuck on running.