Regent Reductions Need Fine-Grained Atomics When Using Simultaneous Coherence

[lightsighter]

Currently Regent will happily generate normal reductions for simultaneous coherence on the CPU which is actually unsafe. It really needs to generate fine-grained read-modify-write atomic updates when using simultaneous coherence.

Note this is not an issue for exclusive and atomic coherence (at the moment) because Legion provides coarse-grained access to the physical instance for the lifetime of the task for those coherence modes.