ke,ex: rework kernel queue

There was previously a lock inversion in the queue code where we acquired the
queue lock under the thread lock (in the wait_any committing path). This patch
attempts to fix this by instead using an atomic count and deferring the waking
of other threads to a DPC.

src/ex/work.zig

@@ -153,7 +153,7 @@ const Pool = struct {
             .monotonic,
         ) == &self.queue.items.head;
 
-        const active = @atomicLoad(usize, &self.queue.active, .monotonic);
+        const active = self.queue.active.load(.monotonic);
         const threads = self.total_threads.load(.monotonic);
 
         if (threads < max_threads_for_prio(self.prio) and !is_empty and