More misc. rodinia opts

• Make GCM not spill ParallelReduce phis.
• Fix unification for inter-device communication.
• Make notion of parallel indices sets more flexible, and use in infer parallel reduce.
• Misc. schedule tweaks.

juno_samples/cava/src/gpu.sch

@@ -120,6 +120,7 @@ simpl!(fuse4);
 //fork-tile[2, 0, false, true](fuse4@channel_loop);
 //fork-split(fuse4@channel_loop);
 //clean-monoid-reduces(fuse4);
+unforkify(fuse4@channel_loop);
 
 no-memset(fuse5@res1);
 no-memset(fuse5@res2);