Serial BoundPenalty class

[nikohansen]

To have a wrapper like in #265, a new BoundPenalty2.repair function should immediately update the penalty weights. This introduces a slight noise for (change of) the gradient perpendicular to the boundary, but this may not necessarily be a disadvantage.

Conceptual draft of the usecase of BoundPenalty2:

class WrapperBoundPenalty:
    def __init__(self, fun, bounds, size=222):
        self.fun = fun
        self.bound_penalty = BoundPenalty2(bounds)
        self.transform = self.bound_penalty.repair
        self.fvalues = collections.deque(maxlen=size)
        self.penalties = collections.deque(maxlen=size)
    def __call__(self, x):
        self.fvalues.append(self.fun(self.transform(x)))
        self.penalties.append(self.boundary_penalty.penalty)
        return self.fvalues[-1] + self.penalties[-1]
    @property
    def penalty(self):
        """last penalty"""
        return self.penalties[-1]