Use the SCS box cone API

[stephane-caron]

Use the box cone API of SCS.

• Suggested in Tolerance tuning for inequality constraints bodono/scs-python#36 (comment)
• Discussion in Trying out the box cone API bodono/scs-python#63

Performance

Here is a comparison of performance on my machine for:

1. Using the box cone API
2. Stacking inequalities in the positive cone

Problem	Box cone	Positive cone
Sparse $(n=500)$	4.99 ms ± 3.42 µs	6.33 ms ± 7.64 µs
Small dense	1.09 ms ± 9.7 µs	612 µs ± 5.36 µs

On the random dense benchmark the box API does not seem to scale as well though:

Here $bl=-0.3$ and $bu=0.3$ for every coordinate. Here is another variation of the problem where $bl=-0.03$ and $bu=0.03$, so that box inequalities are more often constrained in optimal solutions:

From these observations only, it seems using the box cone performs better than stacking in the positive cone on sparse, but worse on dense problems.