feat(scheduling): add soft task and group concurrency

[psteinroe]

Replace hard concurrency slots with database-coordinated soft task and group limits.

Workers count active executions while claiming with FOR UPDATE SKIP LOCKED. Concurrent claim transactions may temporarily overshoot a configured limit; no slot or semaphore rows are maintained. groupConcurrency applies independently within each populated (queue, task, group) scope, while ungrouped work bypasses the group limit.

Claim performance

The claim path now follows the bounded candidate pattern used by pg-boss, Graphile Worker, River, and Oban: it locks an ordered batch before applying task/group window ranking, instead of ranking the entire ready backlog. It materializes active executions once, skips saturated tasks and groups before locking, and uses a database-derived no-limit fast path so rolling workers cannot bypass newly configured limits. A covering partial index supports both task and group active counts.

With 1,000,000 ready executions and a batch size of 50, median claim latency dropped from 6,138 ms to 45 ms without limits, from 5,570 ms to 48 ms with a saturated group, and from 5,333 ms to 35 ms with a saturated task. A hot group can intentionally underfill one claim because ranking is bounded; the next poll excludes the now-saturated group and continues through eligible work.

Related consistency changes:

• stale or unregistered orchestrators in the in-memory client release normal executions and fail cancelled executions;
• settlement identifies a claimed execution by (id, queue, locked_by) without requiring the duplicated task key;
• cron conflict updates refresh schedule payload, run time, cron expression, and group metadata.

Closes #12