Slow reverse windowing in KMeansAD

[wjyzxcv]

The _custom_reverse_windowing method of the KMeansAD class is very slow for long time series, probably because of a for loop being used when trying to "only iterate over window intersections". It seems to me a simple moving average would be much faster without degrading performance. I know it may not be good practice to change the algorithm for benchmark purposes but I hope you could make it an option.

[qhliu26]

Thank you for your suggestion! We will look into this issue.