Replace p90 and P2Estimator with Simple Exponential Smoother (SES)

[kris-gaudel]

Using the p90 as our ideal error rate is not resilient to 30min incidents. This PR replaces the p90 as calculated by the P2 estimator with the forecast from a SES.

A small alpha is chosen to make reduce the weight on recent events. As our number of observations increases, we become more "confident" in our observations so we reduce alpha by half after 15mins, and half again at 30mins (this idea was proposed by @AbdulRahmanAlHamali, exact numbers and reduction facts are to be discussed).

Some things that are still being considered:

• What should is a good cap value for our forecast?
• What is a good initial value?
• It would be nice to verify these changes on SFR, perhaps our p90 was too sensitive and that's why our utilization spiked here in Adaptive - Test 3