The warpspeed scan implementation uses size_t as offset type in all cases, so there is no need to benchmark or tune I32 and I64 offset types.
The old implementation did distinguish but it's only used as fallback for corner cases now, so tiny regressions of using I64 should not matter.