tp: combine batches which mix implicit and stored ids

BatchBuffer carried its own copy of RowStore's check that a column keeps
its representation from batch to batch, with the same gap: a stored
Uint32 id followed by an implicit one was accepted, the reverse was not.
TreeParentFirst produces the reverse, and its small batches are exactly
what a throughput consumer asks to have combined, so the pipeline failed
here before the rows ever reached the store.

Use the shared comparison instead. Columns which differ only in how they
hold their ids fall through to the packing path, which already writes
both forms out as Uint32.