Use separate cutoff for bonded and non-bonded interactions

[fweik]

Because espresso finds bond partners by an index, and not by spatial neighbor search it is sufficient
for bond partners to be present (real or ghost) on the node where a bond is evaluated. Hence the cutoff for bonds needs only to be considered for the halo size, not for the cell size. This can lead to
considerable performance increase: e.g. currently adding a dihedral bond doubles the cutoff, and hence the cell size. This causes an 8-fold increase in the number of pairs that are considered in the short range pair finding. It would actually be sufficient die just use a two-cell halo and leave the cell size as is, which would not affect the performance of the hot loop.

This is also required to be able to handle multi-body bonds without the need of a reduction after the force calculation, because by selecting an appropriate halo size for the bond types present (1, 2 or 3) allows for evaluating bonds on ghost particles without influencing the performance of the pair loop.

[fweik]

Furthermore the bonded cutoff can then be ignored on a single node and for the n^2 cell system, because in either case all particles are always present.

[jngrad]

Several points have already been addressed in the upcoming 4.2.0 release, while the cell size remains an issue:

• the bonded cutoff and non-bonded cutoff are now treated separately in the short-range loop
  

  espresso/src/core/short_range_loop.hpp

  Lines 47 to 53 in ef9f497

  	if (bond_cutoff >= 0.) {
  	cell_structure.bond_loop(bond_kernel);
  	}
  	if (pair_cutoff > 0.) {
  	cell_structure.non_bonded_loop(pair_kernel, verlet_criterion);
  	}

• the bonded cutoff is now ignored on 1 MPI rank
  

  espresso/src/core/interactions.cpp

  Lines 59 to 63 in ef9f497

  	if (not single_node) {
  	// If there is just one node, the bonded cutoff can be omitted
  	// because bond partners are always on the local node.
  	max_cut = std::max(max_cut, max_cut_bonded);
  	}

• the bonded cutoff is still used for the skin, and hence for the cell size via mpi_set_skin() and cells_re_init()
  

  espresso/src/core/integrate.cpp

  Lines 448 to 451 in ef9f497

  	auto const new_skin =
  	std::min(0.4 * max_cut,
  	*boost::min_element(cell_structure.max_cutoff()) - max_cut);
  	mpi_set_skin(new_skin);