Correctly update atomics in IoDriverMetrics

[asomers]

Updating an atomic variable with a load followed by a store is racy. It
defeats the entire purpose of using atomic variables, and can result in
"lost" updates. Instead, use fetch_add .

Motivation

Avoid inaccurate metrics due to update races.

Solution

Always use atomic instructions to update atomic variables.

[asomers]

I think the failure on windows-latest is spurious. The --unstable test on windows-latest worked, after all.

[Darksonn]

Thanks. I'm surprised this wasn't caught in the PR originally adding this.

[vitalyd]

(just lurking) The original code might've been intentional since fetch_add(1, Relaxed) is going to issue a lock'd instruction on x86/64 (godbolt). An argument could be made (I wouldn't make it without some benchmarks to support this, but ...) that for stats/metrics, the chance of a lost update is the price to pay to generate lighter weight code (i.e. the original code is a normal store, no memory fences).

[Darksonn]

@vitalyd The original argument for doing it appears to be that only the IO driver changes these counts, so there isn't actually any race. However, looking at the code it does not seem like that is the case. The incr_fd_count method can be called in parallel for example.