Three optimizations to reduce SQL round-trips per task lifecycle:

1. peek_next uses subquery so the partial index on (status, priority,
   id) acts as a covering index — the outer SELECT fetches only the
   winning row by primary key instead of scanning all pending rows.

2. Add complete_with_record / fail_with_record that accept the
   in-memory TaskRecord, skipping the redundant SELECT * that re-read
   a row the caller already held. Dispatch and parent-resolution paths
   now use these variants. The requeue flag (which may be set by a
   concurrent submit) is handled via conditional DELETE/UPDATE instead
   of a SELECT-then-branch.

3. Replace compute_duration_ms SQL (julianday round-trip) with pure
   Rust chrono arithmetic — one fewer statement per transaction.

Net effect: complete() goes from 5 SQL statements to 3, reducing
write-lock hold time. Benchmarks show 9-16% improvement on
concurrency_scaling and 7% on mixed_priority_dispatch.

…into focused submodules

Break down the three largest files by separation of concerns:

- scheduler/mod.rs (1440→199 lines): extract submit.rs, run_loop.rs,
  control.rs, queries.rs, and tests.rs
- registry.rs (537 lines): convert to registry/ directory with context.rs,
  state.rs, child_spawner.rs, and io_tracker.rs
- task.rs (751 lines): convert to task/ directory with submission.rs,
  typed.rs, error.rs, dedup.rs, and tests.rs

Also deduplicate SpawnContext construction via build_spawn_context() helper.
No public API changes — all types remain re-exported at the same paths.