Improve default nearest resampling radius of influence

[djhoese]

See #521. The KDTree-based resampler in Satpy is powered by pykdtree. If the user doesn't force the value for radius_of_influence (the geocentric distance to look for valid pixels until giving up and putting in a fill value), then pyresample checks the "geocentric resolution" estimates of the input area and output area and takes the maximum between the two. In the AMSR2 high resolution band case that means the 5000m hard-coded nadir resolution is multiplied by 2 to get an estimate for the resolution at the limb of the swath. For most high resolution output grids this value of 10000m will always be larger and therefore be the place where work should be done.

In pyresample/geometry.py there is a SwathDefinition.geocentric_resolution method. This method should be updated to check for a limb_resolution attribute and use that to make a better estimate of the limb resolution. Either the maximum of the nadir and limb or an average (could cause holes). The amsr2_l1b reader in Satpy should then also be updated to include a limb_resolution attribute for the updated pyresample code to use.