Priority Level

Low

Task Summary

Two related cleanup gaps in the async scheduler's row-group lifecycle, found during #429 review. Neither is a correctness bug - both are bounded and low-risk - but they're structural patterns that will compound as the scheduler evolves.

Technical Details & Implementation Plan

1. RowGroupBufferManager - add a "discard without write" path

checkpoint_row_group() is the only method that frees _buffers[rg_id], _dropped[rg_id], and _row_group_sizes[rg_id], coupling memory cleanup with parquet I/O. Any failure path that skips the write (e.g. on_before_checkpoint raising) leaks the buffer until _build_async returns.

Fix: add free_row_group(rg_id) that does just the del/pop cleanup. checkpoint_row_group calls it after writing. The dropped branch in _checkpoint_completed_row_groups calls it directly.

2. Fold _seeds_dispatched_rgs / _pre_batch_done_rgs into _RowGroupState

These per-RG lifecycle markers are standalone sets that require explicit .discard() on completion - the same class of bug that already hit _admitted_rg_ids. Moving them into _RowGroupState as fields means del self._rg_states[rg_id] handles all cleanup automatically.

Dependencies

Follows #429