Check if per-matrix sparsity (automatic) makes sense

[greschd]

Currently, the hopping matrices can either all be stored as dense matrices, or all sparse. It might make sense to mix sparse and dense matrices though.

Since there's anyway a compatibility layer in the form of _array_cast and empty_matrix, it might not be too hard to implement.

What's more difficult is probably adjusting the sparsity dynamically. A good starting point could be analyzing the matrices in set_sparse, and then set each according to some heuristic for the filling factor (adjustable?).

To be checked:

• Find some models where mixed sparse / dense provides measurable benefit.
• Find cut-off where storing dense / sparse is more efficient
• Check impact on Hamiltonian evaluation
• Are there places where matrix manipulation relies on it all being the same data type?