2.0 backfill tolerance update

[jranson]

This patch changes the backfill tolerance pattern, which is used to prevent preliminary real-time data from being cached until it is old enough to be considered final, from an eager load to a lazy load pattern.

The previous implementation would simply not cache data new enough to fall in the backfill tolerance window. As a result, every single real-time request to Trickster for the given Backend would result in a partial cache hit at best, since some data would always be excluded from the cache and need to be fetched. There was no possibility of a Cache Hit for time series requests on backends with a backfill tolerance set.

The new approach improves the methodology by caching the the preliminary data and serving to subsequent requests, so long as the lookup is otherwise a cache hit. once a request is made that has a cache status of partial hit - meaning some data requested by the client, but in a different time range than the preliminary data, needs to be fetched - then the resulting delta proxy request will also include any preliminary data ranges that were part of the client range, so as to refresh them.

In this way, preliminary real-time data are eventually updated to their final values, so long as subsequent real-time requests continue to come through to create qualifying partial hits. This allows backends with backfill tolerance to have a Cache Hit Rate comparable to Backends without a backfill tolerance set, while ensuring preliminary data still gets updated.

Separately, a new option was added called BackfillTolerancePoints (in addition to the existing BackfillToleranceMS). The new option will calculate an alternate backfill tolerance of BackfillTolerancePoints * request.StepValue, and the delta proxy cache will choose whichever tolerance value is greater, based on the query's step value. This way, no matter if a chart is a 1yr look with 1wk steps, or 10m look with a 5s step, backfill tolerance will work proportionately.