This is a follow-up to #15702 that hopefully claws back the 1%
performance regression. Assuming it works, the trick is to iterate over
the constraints vectors via mut reference (aka a single pointer), so
that we're not copying `BitSet`s into and out of the zip tuples as we
iterate. We use `std::mem::take` as a poor-man's move constructor only
at the very end, when we're ready to emplace it into the result. (C++
idioms intended! 😄)
With local testing via hyperfine, I'm seeing this be 1-3% faster than
`main` most of the time — though a small number of runs (1 in 10,
maybe?) are a wash or have `main` faster. Codspeed reports a 2%
gain.
5a9d71a5f1