with sparse and hybrid systems, Pair::ev_tally() may not be called on
every processor and thus the computes in USER-TALLY may hang during
reverse communication because of the error->all() call after checking
whether callback from Pair::ev_tally() has been called at least once.
To address this cleanly, a second callback function needs to be added,
which is run during Pair::ev_setup() and will now handle all memory
re-allocation and clearing of accumulators, just like it is done for
regular tallied data.