Fix ThreadCachedInt race condition

Fix ThreadCachedInt race condition

Summary:
Acquire a SharedMutex at ThreadExit to ensure that after unlinking the ThreadEntry from
the list, future accessAllThreads() won't miss a destroying thread.

This is quite a dangerous fix as it changes some lock ordering semantics. ThreadLocal
elements are now destroyed while holding a lock, so if the destruction function
acquires a different lock, ordering must be consistent with other
uses of accessAllThreads().

I've made accessAllThreads() an opt-in feature via a template parameter and changed
all existing uses. I've also fixed a few lock ordering issues that arose due to this
change.

Reviewed By: andriigrynenko

Differential Revision: D3931072

fbshipit-source-id: 4d464408713184080079698df453b95873bb1a6c