error of the first interval of the IntegratorLearner
Since efcccfcc we set the error of the very first interval to a very large number (largest Python float):
ival.err = sys.float_info.max
we do this because when we split the interval (and we don't know the function values yet) we want the error to go down, and inf/2==inf
so that doesn't work.
To quote @anton-akhmerov:
We're now postulating that as long as there's a single unevaluated interval, it has the highest priority compared to any other. This means that if we request 200 points, provide 100 back, request 100 more, etc, we're always going to stick with evaluating a random ever shrinking incompletely evaluated interval.