Remove ODEProblem

[SebastianM-C]

This PR removes the ODEProblem code path since it was not needed
and it made algorithm support for run_simulation unnecesarly complicated (see #26).
Also, the ODEProblem code did not support thermostats from what I can tell.

Since the ODEProblem dispatch was used for the default algorithm choice with Tsit5
as the default, this changes to VelocityVerlet with a timestep equal to tspan/10000.
I opted for VelocityVerlet since it is a very popular algorithm for N-body problems.
The disadvantage compared with the previous default is that it increases runtimes
with about ~2 orders of magnitude, but the energy and other invariants of the
system should be conserved better. Maybe after the benchmarks we can choose a
better default.

I also adapted the indexing in some tests to match with the new solution type
and added some comments explaining the tests (what we test from the physics point of view).

[ChrisRackauckas]

Note Tsit5 works on DynamicalODEProblem, so it can still be used, but indeed it shouldn't be default in these cases. The default should be chosen based on benchmarks.

However, I don't think a default dt makes sense if an algorithm isn't adaptive.

[SebastianM-C]

I put the dt because otherwise we can't have a simple run_simulation(simulation) method as shown in the README. Should we use an adaptive algorithm by default in this case?

[ChrisRackauckas]

I think dt should be a required keyword argument if we use a non-adaptive default.