using the return values of HashSet::remove(), we can factor
pairs of contains()/remove() into a single remove().
On a perfdiscovery run done on the PyPy repository, prepared
with contrib/discovery-helper.sh 50 100, I do get a modest improvement
with this (mean of medians of three runs is better by 2%)
Sample readings, before this change:
! wall 0.175609 comb 0.180000 user 0.180000 sys 0.000000 (median of 58)
With this change:
! wall 0.171662 comb 0.180000 user 0.170000 sys 0.010000 (median of 60)