Use more internal iteration

[frankmcsherry]

A lot of the complexity in DD is "external iteration", where abstractions need to reveal their internal types to support a generic algorithm. This is "anti abstraction" where detail spills outwards, and specific implementations are disempowered. This seems to happen in a few places, that we're going to try to clean up in order:

1. Update container consolidation, as part of chunking.
2. Merging in the merge batcher.
3. Building into trace batches.

The first step seems pretty easy, and a reduction in code. It adds a new container trait that consolidates from one instances of the container to another. There is the opportunity to capture a common pattern and provide it, but no longer the requirement to do so. The TStack chunker is not yet replaced, in that its behavior is more of a vec chunker that is followed by a conversion to a columnation container, and .. there's probably an easier way to express that.

The second step is also pretty easy, but a bit more new code is written. Rather than externalize merging with container queues and such, there is a trait with two methods,

        fn merge_from(
            &mut self,
            others: &mut [Self],
            positions: &mut [usize],
        );

        fn extract(
            &mut self,
            upper: AntichainRef<Self::TimeOwned>,
            frontier: &mut Antichain<Self::TimeOwned>,
            keep: &mut Self,
            ship: &mut Self,
        );

which allows an internal implementation of the merge batcher. It misses on some performance nits, like extract now extracts from one container into two to completion, rather than popping up for air as either of the destinations "fill". But it's also ~100 lines to put a different merger in place, which seems easier now.

[frankmcsherry]

I think we'll stop without doing the third step. It also seems worth fixing, but it's a moment where external iteration is acting as a m:n bridge between container types that are fundamentally different (containers of updates, and then columns of batch containers). I can imagine eventually wanting to unify these as chains of the same type, but tbd on just how easy/hard it should be when folks want to shift container types.