Revisit the stashing logic in MergeBatcherColumnation

[antiguru]

As surfaced recently, the stashing logic to recycle empty buffers in the merge batcher is subtle; it should make sure there are at most 2 empty buffers, which is all it will ever need. The vector-based merge batcher implements this, but the logic is not simple to understand. When implementing the columnation-based merge batcher, I got it wrong, causing the seal function to temporarily retain all empty buffers, which caused an OOM situation.

We should revisit that the implementation maints the invariant that there are at most two empty buffers, and reason about why this is sufficient.