Make data predicate evaluation column-major

[dominiklohmann]

This is a major performance optimization for expression evaluation that essentially fills in one of the last TODOs from #2440: Making data predicate evaluation itself column-major, i.e., moving as much as possible of the evaluation out of the hot loop.

Here's how the new expression algorithm works, roughly, with (3.ii) and (4) being the new and improved steps:

1. Normalize the selection bitmap from the dense index result to the length of the batch + offset.
2. Determine whether the expression is empty, a connective of some sort, or a predicate. For connectives, resolve them recursively and combine the resulting bitmaps accordingly.
3. Evaluate predicates:
   1. If it's a meta extractor, operate on the batch metadata. In case of a match, the selection bitmap is the result directly.
   2. If it's a data predicate, access the desired array, and lift the resolved types for both sides of the predicate into a compile time context for the column evaluator.
4. The column evaluator has specialization based on the three-tuple of lhs type, relational operator, and rhs view. The generic fall back case iterates over all fields per the selection bitmap to do the evaluation using the cell evaluator, which can be specialized per relational operator.

We've seen some pretty great results. When using a query on a 6M event database that's doing multiple substring searches (for which we have neither sparse nor dense index support), the full scan of the database is now over 2x quicker than before.

Concretely, we measured the following numbers using Hyperfine with three warmup runs and ten measured runs for this commit compared to its merge-base with master de86af3fca.

New:

Time (mean ± σ):     11.038 s ±  0.360 s    [User: 28.600 s, System: 4.430 s]
Range (min … max):   10.808 s … 11.569 s    10 runs

Old:

Time (mean ± σ):     22.805 s ±  0.616 s    [User: 45.060 s, System: 4.826 s]
Range (min … max):   22.042 s … 23.568 s    10 runs

This was a Hackathon group effort with @patszt, @dispanser, and @dominiklohmann.

[mavam]

Looks good at a high level. I only left some minor questions/comments. If CI passes, this would work for me.