Replace the use of std::sync::Once with no_std-compatible code

[briansmith]

See #744. One of the two pain ports in porting ring to a no_std platform is our use of std::sync::Once, which (obviously) requires libstd. There is only one use of std::sync::Once. Our single use of std::sync::Once is very simple: We look up the CPU capabilities (e.g. using CPUID) and store the results.

I think it is safe to assume that looking up the CPU capabilities is idempotent, so we could probably even let threads race with each other to cache the CPU capabilities, letting them stomp on each other's caching (in a controlled way).

/cc @jethrogb

[briansmith]

Another solution would be to use parking_lot in no_std mode.

[briansmith]

Fixed in 0.14 by using spin::Once.

[briansmith]

Fixed in 0.14 by using spin::Once.

This caused #931: CPU feature detection may be racy.