Use HierarchicalMatrices.jl for n-body problems

[dlfivefifty]

This continues a conversation that started here:

SciML/MultiScaleArrays.jl#25 (comment)

HierarchicalMatrices.jl gives a convenient way to implement the fast multipole method, which exploits low rank structure to make matrix multiplication fast. This can be used for efficient simulations: O(n) cost for n bodies.

[ChrisRackauckas]

Right now we have the users defining the potentials:

https://github.com/JuliaDiffEq/DiffEqPhysics.jl/blob/master/test/nbody_test.jl#L22

which isn't a good thing is we want to make this easy and fast. We really should be using a fast multipole method, so we might want to constrain the input potentials a bit so that way we can support this, but I 100% agree that we should do this sense it would make the n-body stuff scale a lot better. In addition, it should probably take into account neighborhood tracking so it's not actually doing every pair of interactions. That's important in MD of macromolecules like proteins.