timour.katchaounov has noticed that for a collection with fields {f_0, f_1, f_2, f_3, f_4}, a collection scan over all the documents is 1.5x - 2x faster when each field is a small string compared to when each field is an integer.

The goal of this ticket is to investigate this, and if it's confirmed to be a real performance difference, determine why the integer comparisons are slower. We should test both the classic engine and SBE to determine if the regression is engine-specific, or something more fundamental to the execution system.