Adaptive threshold for streaming scores (v3.3 backlog item 10)

Hi, I'd also like to pick up item 10 from the same backlog, the adaptive thresholding layer.

The idea is a wrapper around a fitted detector that keeps recalibrating the threshold from a sliding window of recent scores. I'd start with two options: DSPOT (Siffer et al., "Anomaly Detection in Streams with Extreme Value Theory", KDD 2017) and a simple rolling quantile. It wouldn't change BaseDetector, and it only needs numpy/scipy.

Same as #743, happy to wait until #716 and #717 are through.