Create less intermediate data from events

[thorstenhater]

This PR optimizes event handling by removing the intermediate structure created by this chain
of transformations:

1. event_lanes := vector<vector<event>> (one lane per local cell, each lane sorted by time)
2. staged_events_per_mech_id := vector<vector<vector<deliverable_event>>> (one vector per mech id and time step, sorted by time)
3. vector<event_stream> one stream per mech id

The following optimisations where performed:

• cut out the middle step (2) completely as it is wholly unneeded and sort directly into event streams
• remove a spurious index structure from cable_cell_group
• slim down deliverable_event and deliverable_event_data
• event_stream now uses a partition instead of a vector of ranges for splitting its data into dt buckets. (Save 8B per dt ;))

The result is that the quite pathological example calcium_stdp.py (as given in which generates immense amounts
of spikes using a single cell group and a single epoch drops from 3.8GB heap to 1.9GB heap usage at peak
using the same runtime or slightly less (<5% difference).

TODO

• Port to GPU.
• Tests pass
• Examples run through
• Fix tests... Interestingly locally all tests pass on my dev machine regardless of optimisation, vectorisation, and assertion settings.

[boeschf]

cscs-ci run daint-mc

[boeschf]

looks good, some code is duplicated between the backends, see this PR.