Summary:
lock bts and lock btr with the varying lengths turn out to be faster than
a single CAS instruction

fetch_set uses lock bts and fetch_reset uses lock btr when possible.
std::atomic::fetch_or and std::atomic::fetch_and with a mask otherwise

Reviewed By: yfeldblum

Differential Revision: D9652759

fbshipit-source-id: e6d10fdb119a10e15e32131c7389f1d4203a0d91