Use relaxed atomics

[Amanieu]

This PR fixes #8 by using relaxed atomics based on this paper: http://www.di.ens.fr/~zappa/readings/ppopp13.pdf

The implementation is based on this one for C++11: https://github.com/Amanieu/asyncplusplus/blob/master/src/work_steal_queue.h

There are several major changes:

• A deque no longer uses a buffer pool, it instead keeps a linked list of all previous buffers when it grows. These buffers are freed when the deque is dropped. This is required because the implementation described in the paper requires that old buffers are not modified after a grow since a thief might still be accessing them. This is an API change, which requires a version bump.
• The code now properly handles integer overflows for the top and bottom indices by using wrapping operations.
• Worker is no longer Sync, since it can only be accessed by one thread at a time.
• Reduced the initial deque size from 128 elements to 32 elements. This should be enough in the majority of cases.
• A lot of code cleanups

[lhecker]

Thanks for taking care of this, @Amanieu! :blush: I think it was quite obvious that you're a lot more experienced in the usage of atomics than me.

P.S.: I'll make sure to catch up to you soon. :smile:

[Amanieu]

@kinghajj Any chance of merging this?