chore(anomaly_detection) Handle timeseries with very low variance

[ram-senth]

This PR fixes two issues with anomaly detection logic with regards to time series with zero variance:

• SuSS logic for window size selection fails
• the interquartile range based scoring logic results in false positives

These two issues are fixed by

• defaulting the window size to 3 if SuSS logic fails
• using mean of a context window instead of IRQ for time series with low variance

It also includes additional changes to increase test coverage of anomaly detection code and additional tags to track frequency of low variance time series.