Avoid a race condition between dlclose and thread termination.

Implement __cxa_thread_atexit and __cxa_thread_atexit_impl. This
functions are used for destructors of thread_local objects.

If a pending destructor exists, prevent unloading of shared objects.
Introduce __dl_cxa_refcount interface for this purpose. When the last
reference is gone and the object has been dlclose'd before, the
unloading is finalized.

Ideally, __cxa_thread_atexit_impl wouldn't exist, but libstdc++ insists
on providing __cxa_thread_atexit as direct wrapper without further
patching.