Improve scheduler reliability: isolate step failures, catch Redis errors in export

[t0kubetsu]

Summary

Fixes #142. Three related reliability gaps in scheduler.py.

Changes

1. Step isolation in _run_scheduler_step

A bare step_func() call with no error handling meant one failing step silently skipped all remaining steps for the tick (report generation, job cleanup, etc.). Wrapped in try/except Exception with logger.exception — each step now fails independently.

2. Failure tracking in task_master_of_puppets

APScheduler silently swallows unhandled tick exceptions. Added an except Exception clause that logs with logger.exception and increments db.app.config["scheduler_failures"], making scheduler health observable.

3. Widen export exception handler

task_export_to_dbs only caught MeilisearchApiError and HTTPError. A Redis error during flush_batch() would propagate uncaught, causing a partial Meilisearch write with no Kvrocks update (stores diverge silently). Added redis.exceptions.RedisError, OSError, and json.JSONDecodeError to the except clause. Changed logger.error to logger.exception to preserve the traceback.

Test plan

• Kill Kvrocks during an export tick — verify other tick steps still run and error is logged
• Verify Meilisearch export error is logged with full traceback (not just the message)
• Verify db.app.config["scheduler_failures"] increments on tick exception