Avoid extrapolation in nearest neighbour interpolation

[jonasbhend]

When truth and forecast data are mapped, forecasts are extrapolated ad infinitum even if the forecast data covers only a subset of the domain. This is currently the case for INCA forecasts and the inner verification domain (aka most of the high-resolution domain). To avoid misinterpretation of evaluation results, we propose to suppress extrapolation and report missing values for regions that are not fully covered. This PR implements this change.

Spatial coverage fraction

To allow for varying number of stations with observations by parameter, we compute the missing value fraction as the number of non-missing forecast / obs pairs divided by the number of non-missing observations. Currently the missing-value-fraction threshold is set to zero, because in this framework, we don't expect any missing forecasts other than incomplete domain coverage, which should indeed be masked.

Summary of changes

• report distance between nearest neighbours
• estimate native grid spacing
• set extrapolated values to missing based on distnace with nearest neighbour and estimated grid spacing
• compute missing value fraction for area averages of verification results and mask results where too many points are missing
• do not mask observation metrics, as in the current implementation observation coverage is 'complete' by definition