Allow concurrent accesses into noskip_mask.

Fixes CDEF portion of #100 (closed). More practically, this allows multiple tile threads at sb64 boundaries (i.e. in the same sb128) to access the same noskip_mask instance without causing race conditions.

(I just learned that you can use operators such as *, ++, |= on atomic integers.)

This is partially a RFC, an alternate solution is to use uint16_t[2] (!225 (merged)), which would then also require some changes to cdef_apply.c. I personally prefer this solution because it's smaller but it's possible some of you don't like the widespread use of atomic_int.