[JJDSNT]

Summary

• make the m68k __sync_lock_test_and_set fallback return the value that was
  in the lock before the exchange;
• remove the now-obsolete m68k warnings from the pthread AutoDocs;
• add a focused test for spin trylock/lock/unlock and pthread_once.

Why

GCC defines __sync_lock_test_and_set as an atomic exchange that returns the
previous contents. The m68k fallback wrote the requested value and returned
that same value. Consequently the first acquisition of a clear spinlock wrote
one and reported one: pthread_spin_trylock() returned EBUSY, while
pthread_spin_lock() kept yielding indefinitely. This is reached by
libstdc++/Mesa through pthread_once() before rendering starts.

The fix preserves the old value while interrupts are disabled and returns it
after the exchange. It does not change the existing m68k locking model. The
separate m68k compare-and-swap fallback is outside this PR.

This is one commit directly on AROS master and has no PR dependencies.

Validation

• complete libpthread.a builds for linux-x86_64 and amiga-m68k;
• the new test builds and links with -Wall -Wextra -Werror on both targets;
• the test passes inside Linux-hosted AROS;
• the linked m68k test binary was inspected: it loads the old lock word before
  the store and returns that saved value;
• cppcheck warning/performance/portability passes for the new test;
• git diff --check passes.

The m68k test binary was cross-built but was not executed on m68k hardware or
an emulator during this validation.