Explanation is mostly in the title.

The bugs in question are these:

1. mod in C with a negative number will return a negative number, which was not what I expected. I replaced naive calls to mod in the deque code with a macro that works the way I wanted (the way that Python behaves).
2. The code to push and pop from the global queue didn't correctly check whether it could in fact pull/push as much as it wanted to from the queue.
3. The call to pthread_mutex_trylock wasn't checked properly--it returns 0 on success, but the conditional treated it as if it returned 1 when locking succeed.

There remain memory errors; valgrind keeps telling me things about reads in freed blocks, which probably means we're missing some objects in our mark. I'm implementing a testing framework to compare against the single threaded implementation to work on this.