fix: partition evaluator thread safety

[skellys]

I found a thread safety bug where some unwanted partitions would make it past the partition filter. I found and tested this fix on a simple Iceberg table with one schema spec, 5 manifest files (so 5 threads used in the thread pool), and thousands of partitions.

I've also solved this with a thread local storage approach to create fewer _ExpressionEvaluator instances under the hood, but I didn't think that'd be worth the additional complexity, especially with per-thread cleanup.

TODO

• discuss approach on how to test this in the test suite itself.

[rustyconover]

Looks great to me Sean!

[Fokko]

This makes sense to me @skellys That's a great catch, and thanks for working on this 👍