Fix incorrect cpu.sleeper accounting

Under contention for the CPU lock, you could end up in a scenario where
waiters on the cpu.sem would be awoken earlier than they should have been.

The problem crosses two functions. lkl_cpu_get and lkl_cpu_put:

https://github.com/lsds/lkl/blob/master/arch/lkl/kernel/cpu.c#L95
https://github.com/lsds/lkl/blob/master/arch/lkl/kernel/cpu.c#L113

The handling of "cpu.sleepers" is incorrect. "cpu.sleepers" should be
incremented when a thread is waiting to get the cpu.sem so that in lkl_cpu_put,
it can call sem_up to wake a sleeper if any exists.

However what it is currently doing is in lkl_cpu_get is incrementing the
sleepers thread count every time it can't get the cpu lock.

https://github.com/lsds/lkl/blob/master/arch/lkl/kernel/cpu.c#L102

So, if a thread fails to get the lock more than once it will increment the
count of sleeping threads more than once. This means as it stands, in lkl_cpu_put,
that one thread trying to get the lock several times can account for multiple calls
to sem_up. Given that it is awoken each time in lkl_cpu_get, that is wrong.

The correct logic is to decrement the sleepers count after the sem_down returns in
lkl_cpu_get and not in lkl_cpu_put. That will correctly match up cpu.sleepers with
the number of threads waiting on cpu.sem.

arch/lkl/kernel/cpu.c

@@ -103,6 +103,8 @@ int lkl_cpu_get(void)
 		__cpu_try_get_unlock(ret, 0);
 		lkl_ops->sem_down(cpu.sem);
 		ret = __cpu_try_get_lock(0);
+		if (ret > -2)
+			cpu.sleepers--;
 	}
 
 	__cpu_try_get_unlock(ret, 1);
@@ -140,7 +142,6 @@ void lkl_cpu_put(void)
 	}
 
 	if (cpu.sleepers) {
-		cpu.sleepers--;
 		lkl_ops->sem_up(cpu.sem);
 	}