Adaptive schedulers not compatible with particle failures

[odunbar]

Currently schedulers meeting a NaN in the model evaluation will fail when calculating a timestep, even though we may be using a failure handler for the actual update. We should extend this behaviour to the schedulers too

The only current case is (if we are not doing IgnoreFailures) to call the "split_by_success(g)" and pass only the the successful particles into the timestep calculation.