Issue Description:

Cygwin has a custom implementation of swift_once because the compiler only reserves a single word for the "predicate". However, that implementation (in stdlib/public/runtime/CygwinPort.cpp) has a race condition:

• Thread A calls swift_once

• Thread B calls swift_once

• Thread A reads 0 from the predicate

• Thread A stores 1 into the predicate

• Thread B reads 1 from the predicate

• Thread B returns (!!)

• Thread A executes the initialization

There are really three states here: not-yet-run, running, and already-run.