This PR adds the option to scale the cost and constraints of QPs.
In particular, the following options have been added:

• qpscaling_scale_constraints: String in ["NO_OBJECTIVE_SCALING", "OBJECTIVE_GERSHGORIN"]
  Default: "NO_OBJECTIVE_SCALING".
  - NO_OBJECTIVE_SCALING: no scaling of the objective
  - OBJECTIVE_GERSHGORIN: estimate max. abs. eigenvalue using Gershgorin circles as max_abs_eig, then sets the objective scaling factor as obj_factor = min(1.0, qpscaling_ub_max_abs_eig/max_abs_eig)

• qpscaling_scale_constraints
  String in ["NO_CONSTRAINT_SCALING", "INF_NORM"]
  Default: "NO_CONSTRAINT_SCALING".
  - NO_CONSTRAINT_SCALING: no scaling of the constraints
  - INF_NORM: scales each constraint except simple bounds by factor 1.0 / max(inf_norm_coeff, inf_norm_constraint_bound), such that the inf-norm of the constraint coefficients and bounds is <= 1.0.
  Slack penalties are adjusted accordingly to get an equivalent solution.
  First, the cost is scaled, then the constraints.

Implementation details

Minimal computational overhead is associated with the new scaling, in case it is not used. The implementation avoids copying values from the different data structures if they are the same and aliases pointers instead.

Examples & Tests:

• equivalence of primal and dual solutions is tested on different examples including slack variables and one-sided constraints
• some tests were added, showcasing examples which cannot be solved without scaling the QP.

Limitations

• Real-time iterations, i.e. RTI and AS-RTI variants with split preparation and feedback phase are not compatible with QP scaling yet. Work on this has been started at WIP: QP scaling for RTI FreyJo/acados#76 but is not a priority right now.
• Computation of solution sensitivities is not compatible with QP scaling yet.
• Scaling tolerances of the QP solver is not done. Right now, the QP solver tolerances need to be set small enough, for the NLP solver to converge. A follow-up work is described in QP solver tolerance strategies #1553