perf: batch dispatch and completion SQL to reduce round-trips (~56% faster)

[deepjoy]

Summary

• Add pop_next_batch(limit) to TaskStore that claims up to N pending tasks in a single UPDATE…RETURNING…LIMIT statement, replacing the sequential one-at-a-time pop_next() loop when filling concurrency slots
• Add dispatch_pending() fast path in the scheduler run loop that fills all available concurrency slots in one SQL round-trip (falls back to per-task gate-checked dispatch on the slow path)
• Thread a skip_tags flag through insert_history() and complete_inner() so that when the store's has_tags flag is false, the no-op history-tag copy INSERT and delete_task_tags DELETE are skipped entirely (2 SQL round-trips saved per task completion)
• Restructure complete_batch_with_resolve for non-recurring batches to use single batched DELETE FROM tasks WHERE id IN (…), DELETE FROM task_tags, and DELETE FROM task_deps statements instead of per-task loops, reducing SQL from ~5N to N+4 per batch

Benchmark results

Benchmark	Before	After	Change
dispatch_and_complete_1000	~450 ms	~200 ms	-56%
concurrency_scaling/1	~310 ms	~146 ms	-50%
concurrency_scaling/2	~227 ms	~113 ms	-51%
concurrency_scaling/4	~240 ms	~95 ms	-59%
concurrency_scaling/8	~274 ms	~88 ms	-68%
mixed_priority_dispatch_500	~241 ms	~93 ms	-60%
submit_1000_tasks	~60 ms	~60 ms	no change
batch_submit_1000	~13 ms	~13 ms	no change