reduce default executor spins from 4 to 0

[tzcnt]

Previously the work stealing loop would only pause between each full round of work stealing attempts. Now, there is a pause inside of the work stealing loop to reduce cross-cache traffic. As a result, the overall work stealing operation takes a bit longer to complete. That change went out in v1.6. BTW, I profiled it across all workloads and didn't observe any negative effects - it seemed to only help.

I have, however, noticed that extra spinning doesn't seem to be helpful under this new regime. It pollutes lightly-loaded / latency-bound profiles with a lot of ex_cpu::make_worker time that isn't doing anything. For cases where you want to actively spin to minimize latency, you'll want to set_spins() to a much higher than the old default of 4 - probably to something like 100+. So I'm changing the default spins to 0.

There's still 1 guaranteed spin baked into the design - after failing to find work and marking the thread as going to sleep, we have to double check all the queues. Then, there is a std::atomic::wait() on the sleep_wait variable, which does some spinning internally.