Complete the ld.so boot cleanup: move most libc initialization from
_csu_finish() to _libc_preinit(), which is an .init_array function
in shared libc (and mark it INITFIRST) or a .preinit_array function
in static libc, grabbing the _dl_cb callback there from ld.so. Then
in _csu_finish(), invoke the dl_clean_boot() callback to free ld.so's
startup bits before main() gets control.

Other cleanups this permits:
- move various startup structures into .data.rel.ro
- the dl* stubs in libc can be strong and call the callbacks provided
via _dl_cb
- no longer need to conditionalize dlctl() calls on presence of _DYNAMIC

original concept and full diff ok kettenis@
ok deraadt@

Move __cxa_thread_atexit* to its own .c file to avoid pulling the code
(w/ _dlctl reference) into static executables. It's all Mark's code so
put his preferred copyright on it.

ok kettenis@