libraries/base/tests/AtomicModifyIORef is quite slow with N>1

This test case seems to be quite slow for me. It spins up 10 threads and for each one it increments an IORef 10,000,000 times.

This takes 6 seconds to run on my computer with -N1. But once we introduce contention by setting at least -N2, it slows down a lot and runs in the order of minutes(!). Perhaps this is partly some sort of CPU thing. I imagine some CPUs will be better at this than others.

Maybe we can remove a few zeros from the increment count?

I noticed this while running some tests with the nonmoving_thr way, but it looks like this is just the case whenever we have multiple capabilities.