Better OSQP's initialization with active set of constraints detection

## The problem (issue #342)

Following the discussion of #342 it has been noticed that the default zero initialization for dual variables was harmful for convergence, specially when primal variables are feasible and close from the optimum.

## The proposal

From an initial primal and feasible solution, the active constraints can be "guessed". Since the problem is convex, any initialization can be used. Hence, it makes sense to compute un approximation of dual variables that leverage information on the primal variables.

Te idea would be to detect active constraints, for example with `u-Ax