1. Add unit testing for crazier things like random invertible matrices.

2. Add real docstrings everywhere.

3. Try to eliminate the code in matrices.py.

4. Make it work on a cartesian product of cones in the correct order.

5. Make it work on a cartesian product of cones in the wrong order
   (apply a perm utation before/after).

6. Rename all of my variables so that they don't conflict with CVXOPT.
   Maybe x -> xi and y -> gamma in my paper, if that works out.

7. Make sure we have the dimensions of the PSD cone correct.

8. Use a positive tolerance when comparing floating point numbers.

9. Come up with a fast heuristic (like making nu huge and taking e1 as
   our point) that finds a primal feasible point.