Fix acquisition of nested write compat rtld locks.

Obtaining compat rtld lock in write mode sets process signal mask to
block all signals. Previous mask is stored in the global variable
oldsigmask. If a lock is write-locked while another lock is already
write-locked, oldsigmask is overwritten by the total mask and on the
last unlock, all signals except traps appear to be blocked.

Fix this by counting the write-lock nested level, and only storing to
oldsigmask/restoring from it at the outermost level.

Masking signals disables involuntary preemption for libc_r, and there
could be no voluntary context switches in the locked code
(dl_iterate_phdr(3) keeps a lock around user callback, but it was
added long after libc_r was renounced). Due to this, remembering the
level in the global variable after the lock is obtained should be
safe, because no two libc_r threads can acquire different write locks
in parallel.

PR: 215826
Reported by: kami
Tested by: yamagi@yamagi.org (previous version)
To be reviewed by: kan
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks