Optimize core loop of _Count_vbool

[localspook]

To handle the difference between counting ones and counting zeros, _Count_vbool's core loop conditionally complements every block it processes: that's O(n) work. We can do better: we can unconditionally count ones, then, after the loop, if we actually wanted zeros, subtract the number of ones from the number of bits we processed: O(1) work.

[localspook]

All CI jobs are failing with "The agent did not connect within the alloted time of 45 minute(s)." 🤔

[StephanTLavavej]

If that happens, ping me on #stalled-checks on the STL Discord and I can rerun the checks. Unfortunately, while our CI is highly reliable when it runs, we're at the mercy of Azure deciding to give us any VMs, and it's been rough lately.

[AlexGuteniev]

Can you post benchmark results before / after?
(I guess there is such benchmark already)

[AlexGuteniev]

I mean intuitively it looks like an improvement, but without benchmarking I wouldn't be sure that the compiler doesn't do something clever on its own.

This PR optimization is not what compiler is likely to do, but it might be able to achieve the same by duplication the loop for *_VbFirst and ~*_VbFirst and making the condition out of loop.

I don't think it is very likely to happen, and that's why we implemented the _Select_popcount_impl thing, but would be good to prove with measurement.