Interpolate nans linearly

[reneburghardt]

Pollys comment:

we probably want not just replace nans with the mean values but sometimes actually interpolate nans linearly, within certain window (like if there are no more then 3 nan values in a row - they should be interpolated linearly) and only if the nans overflow the tolerance window - replace it with local means

[reneburghardt]

Some concerns of Fabian:

I would not come up with something complicated to fill the NaN values. Either they are returned, the times should be set to not valid, or the user can provide a fill value that replaces NaNs. If we adaptively search around the NaN values weird effects can happen. For instance, imagine we have a spike, then half a second NaN, then a spike. This would result in a half-second long spike. In addition it makes everything slow.

Tom has once started with an implementation of nan interpolation: schewskone@700fa9b