std.Thread.Futex addition

[kprotty]

Having a futex primitive is the first step towards building unified/simpler/efficient synchronization primitives. Some notes on platform compatibility:

• Linux already has a futex api. If unfamiliar with futexes in general, its a good resource.
• Windows uses the WaitOnAddress functions which are only available on Windows 8 and higher. This kills windows 7 support, but we could add an NtKeyedEvent futex-emulation backend if xp to vista/7 is still needed. The implementation uses the ntdll functions because the kernel32 ones aren't actually in synchronization.lib.
• Darwin uses the new undocumented __ulock_wait/wake API available since 10.12. Its also used in LLVM's libc++/libcxx so it should be still available. There's a possibility it may prevent apps from being uploaded to the AppStore since dynamically linked functions to odd places are checked iirc. If this becomes an issue, we could switch to the generic posix backend.
• The generic posix backend emulates the futex api using a global hash table of wait queues. Its purposely made simple for now but could be optimized by using fibonacci hashing on the address, using a custom Lock for each bucket which can be zero-initialized to reduce hash table size in binary with the .bss segment, and using a balanced tree (red-black, avl, treap, etc.) of queues for each address instead of a flat list. This generic hash table futex approach can also be used for other OS's without a native futex api besides posix.
• For other OS's, we could specialize on their futex-equivalent apis in the future. For reference, WASM is working on one, FreeBSD has one, OpenBSD has one, DragonflyBSD has one, but NetBSD does not meaning it would need futex emulation like the generic posix backend.

[lithdew]

It looks like quite a few platforms only support waking up either 1 or all waiters. What do you think about replacing num_waiters: u32 to wake_all: bool, or replacing wake() with something like notify() (wake one waiter) and broadcast() (wake all waiters)?

[kprotty]

num_waiters is more of a hint and can be used to influence the exact count for platforms that support it. Other platforms would just wake up more threads than stated. The waiters already have to handle spurious wake ups, and this keeps it consistent with linux's futex(2) api.

[andrewrk]

Great work!