Better NaN/error handling

[tilsche]

There might be a reasonable way to handle NaNs:

• Record them as raw values
• But don't include them in any of the aggregates
• Reduce the active time for the interval before the NaN value

[bmario]

After a short discussion, the problem here is that count needs to be dealt with as well, as it is used for indexing AND calculation of the mean value. A possible solution would be a nan_count, but that is a major breaking change. Especially, the file format would change.