Eliminate access to thread_local variables in fast-path (like write barrier)

[EvanKirshenbaum]

[imported from HPE issue 250]

We have a couple of places where thread-local variables are accessed in fast-path. This has its own inefficiencies as two function calls are made to access a thread-local.

The only information that we have in write barrier is address of some local variable. Given that, we can implement a lock-free hash-table like data structure, which can store <page#, ptr> pairs. This can happen on the first construction of a thread-local variable, and be cleared up at destruction time.

For lookup, given any local variable address, we mask out the least-significant 12 bits (4K page), and look it up in the structure. If we don't find the entry, then we can create that in the slow-path.