[ML] Improvements to model robustness to outliers

[tveasey]

This makes changes driven by latest round reviewing 6.5 results against our QA suite. It primarily targets model robustness to outliers.

In particular, it

1. Mitigates the impact of outliers on change point detection,
2. Delays detecting changes for slightly longer (to reduce false positives detecting time shifts),
3. Increases the sample p-value at we apply minimum weight for model update.

[edsavage]

LGTM

I do note the CI error, however this does appear spurious