second round of making per-atom allocations consistent.

now all sections that dimension per-atom local arrays to atom->nmax
will also *test* for atom->nmax and thus we avoid potential segfaults
in all cases where atom->nlocal or atom->nlocal+atom->nghost are 0.
it doesn't make sense to test for these if the arrays are dimensioned
to atom->nmax afterwards anyway.

src/compute_reduce.cpp

@@ -574,7 +574,7 @@ double ComputeReduce::compute_one(int m, int flag)
   // evaluate atom-style variable
 
   } else if (which[m] == VARIABLE) {
-    if (nlocal > maxatom) {
+    if (atom->nmax > maxatom) {
       maxatom = atom->nmax;
       memory->destroy(varatom);
       memory->create(varatom,maxatom,"reduce:varatom");