Narrow phase *flush* multithreaded scaling is pretty bad

[RossNordby]

It's not uncommon to see multithreaded narrow phase flush underperform single threaded narrow phase flush by a bit.

Other heavy stages that dominate in single thread, like the narrow phase itself and solving, scale well and can end up being faster than narrow phase flushing as a result.

Narrow phase flushing is extremely complicated right now, and even incremental changes are nightmarish. It would be nice to rip it apart in favor of a simpler version, if such a thing exists.

For nondeterministic case, you might be able to beat the status quo by just having a worker alongside the narrow phase execution that handles all the new constraints sequentially. Determinism would still require that all the constraint add/removes are known ahead of time to avoid order dependence.

This one's pretty tough, but fairly high value if you can find a way to reasonable scaling. In chaotic scenes on very wide machines, narrow phase flush can be the most expensive part of simulation.