[ML] Anomaly detection fails to identify linear trend

[valeriy42]

In certain cases with low event rate and large bucket span (e.g. bucket_span: 1d with a single event in the bucket), the anomaly detection model fails to recognize the linear trend. As an effect, is also forecasts a decreasing trend smoothly regressing to the mean as the prediction horizon increases.

[valeriy42]

We need to check the parametrization for this edge case. The problem may be related to what we see in #2730 .