Re-factor label-resolution.

[thorstenhater]

Using our standard Brunel example with cable cells instead of LIF cells we notice that
simulation setup take roughly 10x longer than the actual execution. By taking a hard look
at the label resolver functionality, we can reduce the setup time by more than half and reduce
the total runtime by half (was 15s)

Result:

bin/brunel --n-excitatory 10000 --n-inhibitory 2500 --tfinal 10 --in-degree-prop 0.1 --dt 0.01 --lambda 1.5 -g 0.8 --delay 1.0 --weight 0.00010 -G 10 --use-cable-cells
==========================================
  Brunel model miniapp
  - distributed : 1 (mpi)
  - threads     : 8
  - gpus        : no
==========================================


There were 38956 spikes

---- meters -------------------------------------------------------------------------------
meter                         time(s)
-------------------------------------------------------------------------------------------
setup                           0.000
model-init                      5.366
model-simulate                  1.549
meter-total                     6.915

Changes:

• Notice that neither assert-univalent nor round-robin-halt actually have a state and remove those.
• State is now stored in a one-layer hash map with key (gid, tag).
• Remove variant and associated polymorphism, replace by switch.
• Remove the partition from range_set and use linear search.
• Make resolver able to reset, avoiding re-allocation

[thorstenhater]

Merged via #2449