Adjust "loss" definitions

[bstellato]

I would adjust losses to avoid confusions, given the new problem formulation:

1. I would call them risk or violation_risk functions.

2. Make the CVaR the default one instead of having to define it. (We could have a set of usable functions instead of only having cvar...)

3. Define augmented lagrangian inside the package and not necessarily in the loss

[AmitSolomonPrinceton]

Loss is now called lagrangian, as we discussed in our meeting on 08/15 (in the next commit).