Ignore outliers affects scaling but not smoothing

[martin-gorner]

The "ignore outliers" is feature is a great idea. Most training curves start from a nonsensical first point that is best ignored. However, it should also apply to smoothing. In its present implementation, you get correct scale but a "smoothed" curve that still starts at the nonsensical first point and slowly descends towards real data, even if in reality, it had only one initial nonsensical point and descended towards normal values immediately.

An easier heuristic-free implementation would be to just ignore all data points before N. For scaling, smoothing, everything. Also, the first few points are always nonsensical but I usually do care about outliers that happen during training.

[lucasb-eyer]

#610 is somewhat related.