Make more improvements to threads and mappings

- NetBSD should now have faster synchronization
- POSIX barriers may now be shared across processes
- An edge case with memory map tracking has been fixed
- Grand Central Dispatch is no longer used on MacOS ARM64
- POSIX mutexes in normal mode now use futexes across processes

third_party/nsync/mu_semaphore.internal.h

@@ -4,6 +4,20 @@
 #include "third_party/nsync/time.h"
 COSMOPOLITAN_C_START_
 
+/* XNU ulock (used by cosmo futexes) is an internal API, however:
+
+     1. Unlike GCD it's cancelable i.e. can be EINTR'd by signals
+     2. We have no choice but to use ulock for joining threads
+     3. Grand Central Dispatch requires a busy loop workaround
+     4. ulock makes our mutexes use 20% more system time (meh)
+     5. ulock makes our mutexes use 40% less wall time (good)
+     6. ulock makes our mutexes use 64% less user time (woop)
+
+   ulock is an outstanding system call that must be used.
+   gcd is not an acceptable alternative to ulock. */
+
+#define NSYNC_USE_GRAND_CENTRAL 0
+
 bool nsync_mu_semaphore_init_futex(nsync_semaphore *);
 errno_t nsync_mu_semaphore_p_futex(nsync_semaphore *);
 errno_t nsync_mu_semaphore_p_with_deadline_futex(nsync_semaphore *, nsync_time);