Update SCS solver syntax

[matthewware]

Would also like to use Convex.ComplexVaraible but the "Objective cannot be a complex expression"

[matthewware]

Maximal dnorm tests are failing in 4 cases. Calculations are accurate to 1e-4 but not to 1e-5.

[matthewware]

@marcusps seems the dnormcptp problem cast as an min of an operator norm is taking many more iterations to converge over the maximization of the trace (if it converges at all). I was still having failures up to 30_000 for random unitaries. This branch uses dnormcptp2 for ddist. Any objections to this?

Also, I tried to switch to Convex.jl support for complex variables. Unfortunately they only seem to work for constraints and can't be in the objective function: https://github.com/jump-dev/Convex.jl/blob/621c212b092b86bea3a3e14c4468a6120d4591ce/src/problems.jl#L11

[marcusps]

@matthewware Thanks for fixing this. The two formulations are duals (both are described in Section 4 of arxiv:0901.4709), so I am not sure why one would converge but not the other.

It may be worthwhile to look into implementing the SDPs from arxiv:1207.5726, but since Watrous does not discuss the numerical properties of the algorithms, we'd have to assess any potential advantages/disadvantages empirically. Nevermind, that paper discussed the general case, not the specialization to CPTP maps. It would be nice to run a performance analysis between the '09 algorithm for CPTP maps and the '12 algorithm for arbitrary linear maps.