[Minor] Implement active set tracking in the inner coordinate descent loop

[tbenthompson]

This is a leftover task from #92

For narrow problems where we compute a full hessian, the inner CD loop is so fast that this would not be helpful to do. However, for future solvers for wide problems where we can't precompute a full hessian, I suspect that implementing inner loop active set logic would be very helpful!

[tbenthompson]

The result of #424 makes me think that this issue might have a substantial impact in certain narrow cases if someone ever wants to implement it.

This will have a large impact if:

1. The problem is coordinate descent dominated. This is normally only true for gaussian distribution and l1 penalty settings (LASSO).
2. The problem has a large penalty so that many coeffs will be zero.

But, in that situation, I could see it potentially leading to decent performance improvements because the innermost loop of the CD solver will iterate over a much smaller list of coefficients.

I'm leaving this "on hold" because it's not a high priority right now but I just wanted to share my thoughts.