Improve Graph space performance

[fxjung]

Currently, large graph spaces are nowhere near as fast as continuous spaces.
Examples:

Distance between random points

• Euclidean 2D: 4.88 µs
• 10 x 10 grid graph: 4.7 µs
• Manhattan Street network: 132 µs

Time interpolation between random points

• Euclidean 2D: 4.91 µs
• 10 x 10 grid graph: 6.11 µs
• Manhattan Street network: 147 µs

Additionally, caching distances between simulations runs might be sensible, see also #100.

[pmarszal]

I found, that often you also need distances and maybe even shortest paths for certain statistics, e.g. what are possible acceptable requests for a certain state of the system. Pre-computing both distances and Dijkstra-predecessors and being able to just give them to the simulator was quite nice for me.