various bugfixes to reference /tally computes

• remove vector compute to avoid ambiguity when using atom style variables. fold reduction into compute_scalar.
• count pair contributions only if one of the two is one group and the other atom in the other group. groups can overlap (all/all counts all pairs)
• after using reverse communication to collect contributions from ghost atoms, we need to zero out the data from ghosts, or they'll be counted twice with compute reduce.

thanks to loris ercole for demonstrating the issues