Speed up subproblem

[phschiele]

Timings in seconds for the original QP/quad form formulation vs the updated formulation.
The number is the total time spent in the subproblem across the 10 iterations of the CPU version of the deepinvtesting notebook.

Solver	QP	sum_squares
Clarabel	256	400
MOSEK	431*	38
OSQP	94	114

*spent a lot of time in linear dependency checker step

Can speed up some more (guess: ~10-15%) by using parameters in CVXPY.
Could mention that switching the solver can have big performance impact in docs, and make the solver choice configurable.

[fabian-sp]

Adding here some stats for the example_residual with d=256, m=32

OSQP:

The whole time spent in self.SP.solve is around 0.05 sec per iter.
The pure cvxpy solve time (SP.solve_time) is around 0.03 sec per iter.

Clarabel:

The whole time spent in self.SP.solve is around 0.1 sec per iter.
The pure cvxpy solve time (SP.solve_time) is around 0.06 sec per iter.

Main question: can we speed up the time not spent for solving (that is, ca 0.02 sec for OSQP), with Parameters?

[fabian-sp]

After changing the seed, tests now also pass for me locally