Fix (allow) thread local allocations from within pthread TLS destructors

Fix (allow) thread local allocations from within pthread TLS destructors

Prevents the GC_thread_key from being cleared on thread exit until
after the thread has been unregistered by GC_unregister_my_thread_inner.

* include/private/thread_local_alloc.h (GC_remove_specific): Call
pthread_setspecific(key, NULL) if USE_PTHREAD_SPECIFIC; update comment.
* thread_local_alloc.c (reset_thread_key): New static function if
USE_PTHREAD_SPECIFIC); define (as macro) to 0 otherwise.
* thread_local_alloc.c (GC_init_thread_local): Pass reset_thread_key
(instead of 0) to GC_key_create.