opt/bench: improve BenchmarkEndToEnd for INSERTs

In `BenchmarkEndToEnd` we have 3 bench cases where we have INSERT
statements. Previously, we always used the same placeholder values,
which forced us to do TRUNCATE TABLE after _every_ iteration, and that
TRUNCATE was included into the operation time. We recently saw
a supposed regression on this benchmark because the performance of
TRUNCATE has regressed.

In my initial approach I tried simply stopping and starting the timer
around the TRUNCATE, but it made the benchmark extremely long. (Timer
operations require stop-the-world pause, and since the time to perform
TRUNCATE wasn't included into the benchmark time, now every single
iteration seemed very short, so we'd do thousands of iterations with the
default `bench-time=1s`, but truncating the table would make it about
1 minute instead.)

To go around this issue I refactored three INSERT queries to generate
slightly different arguments for each iteration so that we don't get PK
duplicates and then moved the TRUNCATE outside the benchmark loop (and
also excluded it from the timer). Now these benchmark cases truly
measure what they were supposed to.

Release note: None

Author: yuzefovich