Implement exponential search for rank -> docid batch lookup

[waamm]

This PR addresses an almost 2 year old TODO by @PSeitz: instead of scanning doc-by-doc to find which document owns a given value rank, we now bracket the answer exponentially and finish with a monotone binary search. Quick informal timings suggest this is roughly ~5× faster on some workloads.

Tests include cases against the old linear-scan implementation.

Edit: On a 10M-document multivalued column where every doc has 2 values, a single select_batch_in_place call converting 10k evenly spaced row ranks (on an M4 Max):

10M nonempty / 10k ranks exponential
select_batch_single_batch        Avg: 2.2253ms (-3.87%)    Median: 2.2268ms (-3.83%)    [2.1889ms .. 2.2818ms]    
10M nonempty / 10k ranks linear
select_batch_single_batch        Avg: 11.9561ms (-3.13%)    Median: 11.8640ms (-3.45%)    [11.7694ms .. 14.0043ms] 

[fulmicoton]

Are there benchmark confirming this is improving something? Can you paste the results in the PR description?

[waamm]

Hi! I added a description of the benchmark I had in mind (I can also add the code).