This is a cheap way to make capacity calculation work well with
multi-group clusters.

There are two alternatives in implementing this:

- we can split the cluster into groups, run individual group
  allocation, and then try to recombine the groups; but this doesn't
  represent the allocation as it would happen in real life, when
  groups are not filled in succession, and would complicate TSPECS
  ordering
- or add this "hack" which does the allocation look at all node pairs
  (per group) at each step, which makes it correct from the point of
  view of ordering of allocations w.r.t. groups

For now, we take the second choice, with the future possibility of
revisiting the first one.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Michael Hanselmann <hansmi@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>