Add `strict_capacity_limit` block cache option to stress/crash tests

As we learned recently with bugs in index partition pinning, there may be bugs when a subset of reads go through block cache. One way that we might expose such bugs is by adding `strict_capacity_limit` to stress/crash tests to make block insertions fail at essentially arbitrary points. This would also give us confidence that our handling of block cache insertion failures is correct.

This issue requires some investigation. It's unknown whether flush/compaction can succeed in case of a full block cache with `strict_capacity_limit == true`. It's also unclear how to retry a user operation after it fails due to full block cache (maybe `ReadOptions::fill_cache == false`?).