QP: later, NLP: learning alg

I decided to make creating of QP class some time later: I should think twice of QP constructor interface. I mean even for LP there are 3 main approaches for assigning linear constraints:
1) A x <= b, Aeq x = beq (MATLAB, CVXOPT) (implemented in OO)
2) A x {less, equal, greater} b (glpk, lp_solve) (implemented in OO)
3) b_l <= A x <= b_u (TOMLAB) (maybe I will implement the one in OO later)

3rd way is more powerful - it allows to reduce A matrix in a factor up to 2 for some cases, but, on the other hand, creating of such matrices A, b_l, b_u is less convenient, and I guess very small amount of LP solvers is capable of taking benefites from the case.

So, a number of QP assignment ways is even greater.

I have to spend several days for investigating some versions of NLP linearisation algorithm that I have obtained from applied systems analysis institute (ASAI), by dr. Danilin. Unfortunately, I haven't English version so I can't discuss them in my blog or scipy mailing lists.