Call __emutls_unregister_key on dlclose

We want to delay the emutls pthread key deletion to keep
__thread/thread_local variables working as long as possible.

Currently, emutls has its own __attribute__((destructor)) function that
deletes its pthread key. That function runs even on process exit, and it
can run before other destructor functions and before C++ static object
destructors.

Move the destructor function to crtbegin_so.c, where an
__attribute__((destructor)) function is only called on dlclose. Use a
priority 0 destructor, which runs after every destructor with a default or
greater priority value.

__emutls_unregister_key will still run before destructor functions of
DT_NEEDED solibs. It also still leaks memory (both the emutls arrays and
each emutls object).

Bug: b/80453944
Test: manual
Change-Id: I6789bcf168415ab8badf2f64687c6a0136c5c917