Cache FileInfo when replaying WAL

[benjaminwinger]

The WAL replayer is currently opening the destination file every time it updates or inserts a page, which is unnecessary and slows down bulk modifications.

In the benchmarks for #2938 it reduced the cost of replaying the WAL (in tmpfs) from 4.5 seconds to 3.2 seconds (where I'm still copying the hash index through the WAL file on the first try, but on additional copies in future significant amounts of data will still need to go through the WAL file).

I don't particularly like this solution, and I feel like it would be better to group them by file (removing the need for hashing and repeatedly looking up the same file), but I don't know of any way reasonable way of doing that without messing up the order that the records are replayed.

I briefly tried using a std::find with a vector of pairs, which gave more or less the same performance. If we allowed c++23 I'd use std::flat_map, but I'm not convinced that the vector of pairs by itself simplifies anything (no hashing, but messier access).

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 91.98%. Comparing base (4d21128) to head (07ec07f).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #3137   +/-   ##
=======================================
  Coverage   91.97%   91.98%           
=======================================
  Files        1171     1171           
  Lines       44041    44047    +6     
=======================================
+ Hits        40508    40517    +9     
+ Misses       3533     3530    -3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.