Fix exclusion tiles sorting on AMD GCN/CDNA (64 threads per wave) (#4268)

The nonbonded kernel uses USE_NEIGHBOR_LIST (useNeighborList)
so host code also must check it instead of useCutoff.

See also https://github.com/openmm/openmm/issues/3462

platforms/opencl/src/OpenCLNonbondedUtilities.cpp

@@ -224,7 +224,7 @@ void OpenCLNonbondedUtilities::initialize(const System& system) {
     vector<mm_int2> exclusionTilesVec;
     for (set<pair<int, int> >::const_iterator iter = tilesWithExclusions.begin(); iter != tilesWithExclusions.end(); ++iter)
         exclusionTilesVec.push_back(mm_int2(iter->first, iter->second));
-    sort(exclusionTilesVec.begin(), exclusionTilesVec.end(), context.getSIMDWidth() <= 32 || !useCutoff ? compareInt2 : compareInt2LargeSIMD);
+    sort(exclusionTilesVec.begin(), exclusionTilesVec.end(), context.getSIMDWidth() <= 32 || !useNeighborList ? compareInt2 : compareInt2LargeSIMD);
     exclusionTiles.initialize<mm_int2>(context, exclusionTilesVec.size(), "exclusionTiles");
     exclusionTiles.upload(exclusionTilesVec);
     map<pair<int, int>, int> exclusionTileMap;