PERF investigate numba perf for coordinate descent solvers

[tomMoral]

I don't remember why we did not used numba to accelerate the coordinate descent for loop, which is often critical for algorithms performances but this should be investigated.

This issue should only be treated once #70 is done, so we have an easy way to report performance gain compared to other solvers.