Specify an API for defining the scale of point about adaptive

[akhmerov]

We could start from at least unifying the treatment of scaling across learners.

For that we'd define a new class, say _Scale (or Scale, Scaler, etc), which would have the following interface:

On every tell we'd have something like:

if self._scale.update(new_point):
    self._recompute_all_losses

And then we'd pass self._scale.value to every computation of the loss function.