Demo a merge batcher that uses the container builder

[antiguru]

Changes the merge batcher to use a container builder to produce its output. This introduces several interesting behavior and unfortunate code changes:

• The merge batcher becomes simpler by delegating chain formation to a container builder. Instead of replicating the logic from capacity container builders, it can absorb whatever builder we give it. The column builder can happily build 2MiB chunks, while the vector/timely stack builders still aim for 8k allocations.
• To avoid regressing, the push_and_add to push two items and add their diffs needs to be defined on the container builder. This arguably is a smell, but I don't know of an alternative, because we need to call copy_descructured for timely stacks :(
• !! The merge batcher doesn't clear recycled allocations. This is because in columnar, a clear moves a column from any variant to typed, potentially discarding its allocation. Avoiding the clear call might leave some memory sitting around for a bit longer, but the container builders are used to clearing on re-use, so no problems expected on that side.
• I took the liberty and addressed some of the to-dos in the columnar example.

The PR is in the wrong order, the merge batcher changes should go into master and then the columnar builder should pick up the new interface. But, this is how I started so I'm leaving it as-is until we talk about it.

[antiguru]

Superseded by TimelyDataflow#613