[ML] Revisit correction for bucket length when computing probability quantiles normalising results

[tveasey]

Currently, we try and account for the different number of results per time interval for different bucket lengths when computing probability quantiles as part of rate limiting normalised scores.

The assumption we make should result in roughly the same score distribution when modelling pure noise (I need to recheck this). However, it will tend to pull down scores when there are long lasting anomalies in the data for short bucket lengths. This has probably been exacerbated by multi-bucket anomaly detection. In particular, we see cases that the correction has the reverse impact to the one we'd like: producing less consistent scoring for different bucket lengths.

We should revisit whether any sort of bucket length correction rate limiting normalised scores is a good idea.