Priority Level

High

Task Summary

Transform the dataset builder from sequential column-by-column processing into an async task queue with dependency-aware scheduling. Generators become async-first, and the builder dispatches individual cell/batch tasks as soon as their upstream dependencies are satisfied — enabling pipeline parallelism across columns and rows.

Technical Details & Implementation Plan

• Dependency map: built from each column config's required_columns property (Jinja2 template introspection) — no config schema changes needed
• Completion tracker: lightweight columns × rows matrix that determines task readiness
• Async task scheduler: replaces the sequential _run_batch loop; dispatches tasks as dependencies are met, bounded by semaphore
• Generator async migration: all generator types get async-capable agenerate methods (LLM generators already have native async from feat(engine): env-var switch for async-first models experiment #280)
• Row group checkpointing: parquet written when a row group fully completes

Dependencies

Part of #260. Builds on #280 (merged). Related: #269, #344.