ThreadLocal gcc optimization

https://github.com/couchbase/ep-engine/blob/3.0.1/src/threadlocal.h

has only two options, win32 and posix.

while on linux on gcc we have __thread keyword, which is known to be implemented much more efficient than posix approach.

under x86 32/64 access to tls variable is gs: selector indexed and offset in that selector is set by runtime-linker.
so no pthread_getspecific is required at all.
no call, no hash tables jumps, no nothing.

please consider to use it.