Speed up with native sparse matrices

In the discretevalueiteration solver, we saw a 1-2 order of magnitude speedup from using builtin sparse matrices: https://github.com/JuliaPOMDP/DiscreteValueIteration.jl/blob/master/src/sparse.jl

It might also be possible to do this with FIB

Read the referenced DiscreteValueIteration.jl src/sparse.jl implementation, then inspect the FIB solver entry point to determine whether the same sparse-matrix approach applies. Compare the current and sparse representations and benchmark the solver; done means the FIB implementation uses the suitable native sparse matrices and demonstrates a meaningful speedup without changing solver results.

Written by the indexing model from the issue text.