Update to latest sprs-ldl version to show performance improvements

[vbarrielle]

I have performed some micro optimization in the sprs-ldl code (sparsemat/sprs#207), and now the performance is similar to the original LDL implementation in C:

LdlNumeric took 3683ms to solve 3 sparse systems
LdlNumeric (C version of LDL) took 3675ms to solve 3 sparse systems

[vbarrielle]

There have been updates in sprs and sprs-ldl, the main difference is the addition of a fill-in reduction method from the state-of-the art library SuiteSparse. Here are the new benchmark results:

LdlNumeric (rust, NoReduction) took 4567ms to solve 3 sparse systems
L2 diff to ref 0.0000002499301225895644
LdlNumeric (rust, ReverseCuthillMcKee) took 3720ms to solve 3 sparse systems
L2 diff to ref 0.00000002052935494084914
LdlNumeric (rust, CAMDSuiteSparse) took 3085ms to solve 3 sparse systems
L2 diff to ref 0
LdlNumeric (C, NoReduction) took 4253ms to solve 3 sparse systems
L2 diff to ref 0.0000002499301225895644
LdlNumeric (C, ReverseCuthillMcKee) took 3488ms to solve 3 sparse systems
L2 diff to ref 0.00000002052935494084914
LdlNumeric (C, CAMDSuiteSparse) took 2979ms to solve 3 sparse systems
L2 diff to ref 0

This shows this new fill-in reduction helps achieving better performance, we have around 1s per solve. As noted in sparsemat/sprs#199, the best performance would be that of CHOLMOD, at 0.46s per solve, but it's a much more involved decomposition. I think we cannot get better performance out of LDL now, CHOLMOD achieves this performance by being supernodal and exploiting dense structure.

[PTNobel]

Hi! Sorry I've been super offline recently, lots of deadlines and a new semester. Thanks for all your work on this! I'll work on integrating this into the main repo in December at the latest.