Currently, we only benchmark our custom atomic types, but don't
benchmark the standard library types. Add benchmarks for ints and uints
that the standard library supports.
Atomic operations tend to be fast when there's no contention but get a
lot slower with contention, so add tests which run the same stress test
in parallel.
Currently, the stress tests start multiple goroutines, but they all have
their own local atomic, so they are not concurrently modifying the same
values. Return a function which can be run concurrently to modify the
same atomic.
Previously, we were not waiting for the spawned goroutines to end.
Refactor the logic to use a common runStress that runs some function
from multiple goroutines.
