Improvements of energy metric calculation

[flora-hofmann-frequenz]

I realized some unrelated things though:

• We rely on the data being sorted in the sum. This is currently usually the case but it's not guaranteed by the service.
• If we have missing data this can be really off, depending on the cause of the missing data. We currently effectively treat NaNs as zeroes and forward-fill gaps in case we use raw data. This is inconsistent and I think we should just forbid resolution of Nones (although this wouldn't really fill gaps in the current version of the service, but in future only).
• I think the start time of the CumulativeEnergy should be the timestamp of the first sample. The end time I think is fine if we always assume resampled data, otherwise we would need to adjust it too.

Originally posted by @cwasicki in #12 (review)