feat: streamed execution in MERGE

[ion-elgreco]

Description

This adds support for streamed execution in MERGE, with two caveats:

• If CDF is enabled we have to materialize since a stream can only be consumed once. I could make it do this only in streaming mode, but I think it might make sense in general
• Initially MERGE derives statistics about the source to prune the target a bit better and allow for more concurrent writers, the problem here is that deriving statistics about source means we have to consume source. Caching that early on will defeat the purpose of the streamed execution. So I've disabled this part of the early filter construction when streaming mode is on.

In the future we can look into the work of influxdb or lancedb. Especially the work of lancedb is interesting since it can allow for replay of the same stream, which would less stressful on memory when we have cdf enabled in merge. This should work well I think with this PR: #3142

I've actually based some of the work in here on top of this already: #3142, so that has to be merged first. Will also rewrite the commits then since they are messy now (:

[ion-elgreco]

I broke CDF by projecting to early and removing the before state, but need to cache when everything is still there, but we can't cache a dataframe when there are table references