Investigate potential overhead of using pre-compiled queries in non-AOT mode

When doing initial perf investigation of AOT/precompiled query scenarios, we saw that for large number of queries (1000) and a relatively small model (10 entities) precompiled (but non-AOT) took longer than regular (non-precompiled) case. This is unexpected and we should investigate. (perhaps testing was wrong and precompiled queries weren't used despite being generated?)
AOT case is around 10x faster than regular scenario.