hard coded WORST_POSSIBLE_RESULT

[tsinggggg]

WORST_POSSIBLE_RESULT for training loss is hard coded to be 1.0. It makes sense for some evaluation metrics but not all, for example when MSE is used and a ’Crashed' trial would get a training loss of 1. it'd be great if the worst result could depend on the metric.

[mfeurer]

Thank you very much for reporting this issue @tsinggggg. We have fixed it in #845. Please reopen this issue if this fix does not work for you.