If Future.cancel() returns False this does *not* mean that the
cancellation failed, so we should not check for this. Instead
we just cancel all the outstanding futures and wait for them to
die.

If the runner instantiates its own executor, then it is also
responsible for calling 'shutdown'. This was the cause of a
bug in the notebook integration where kernels would fail to die
if ProcessPoolExecutors were not properly shut down. This is probably
a bug in IPython, but we fix our usage here also.

Better loss improvement

See merge request !6

Implement BalancingLearner and fix several bugs in 1DLearner.
Also add an abstractmethod `loss_improvement` to learners, that
allows the BalancingLearner to select which learner will give
it the best improvement.

Closes #10 and #13

See merge request !4

This should be an abstract method so that any learner can be used with
the BalancingLearner.  Also add a dummy implementation to the 0D
learner.

Methods with names like "get_losses" imply that they modify no
state, whereas these methods actually modify the state of some of
their arguments.

Dictionary lookups are cheap, so we should not complicated the
code in order to avoid them.

We no longer need the extra "state", we just work it out by using
(cheap) dictionary lookups.

enable runners to be able to "log" the sequence of method
calls on their learner. We can then replay these method calls
on a "fresh" learner for testing or debugging.

See merge request !5

The motivation for making this a static method was to make the
statefulness of '_run' explicit, but this gets cumbersome when we
want to add more state (have to add more parameters). Luckily Python
has a built-in mechanism for dealing with this: regular methods and
the 'self' parameter!