Rhaas/interpolate2 nans

Interpolate2: be more verbose when not being able to find vector comps

Interpolate2: do not assume that all tensorweights are either 0 or 1

Interpolate2: use NaN as a proper initializer rather than assigning it

Interpolate2: remove assert()s from inner loop unless in DEBUG

Interpolate2: support grid functions containing NaN values

while not nice, these are legitimate values that may occur on the grid

Interpolate: initialize tensorweight with NAN