WIP: Make BQPD matrix-free

[cvanaret]

BQPD only uses Hessian-vector products. Therefore, we can use a linear operator $v \rightarrow H.v$ instead of forming the explicit Hessian. Such an operator is available in ASL (AMPL Solver Library). It will also be very convenient for quasi-Newton methods.

Steps to complete:

• transfer the gdotx function from Fortran to C++.
• hide pointers to the OptimizationProblem and current Multipliers in the BQPD workspace (int* lws), then retrieve them in the gdotx function.
• perform the Hessian-vector product within gdotx.
• wrap the OptimizationProblem, the current Multipliers and the HessianModel in a new object (e.g. InequalityConstrainedProblem), and let the subproblem solver handle the function evaluations.

cc @worc4021