Infinite cutoff extrapolation can return NaN values if crop=True

[robbibt]

For our Australian applications, we often run our modelling with both extrapolate=True and cutoff=np.inf to ensure we get at least some modelled tides for areas just outside the modelling domain (even if they're less accurate). I've recently started running code using crop=True, as it makes a huge difference to run-times. However, I've noticed that I'm now getting many NaN values in areas where we used to get modelled tides.

I'm pretty sure this is because the cropping step is resulting in no valid consituent data within the clipped analysis extent, so there's no data to actually extrapolate. I'm not sure there's really a good solution to this - when using cropped model files, there's no way to access data outside the clipped extent without loading it into memory (which defeats the purpose of crop=True). I think there could be two ways to make this a bit less unexpected:

1. Raise a warning if crop=True and extrapolate=True, yet no data is availble to extrapolate. This would at least give the user a heads-up that extrapolation is not performing as expected
2. Expose the buffer used under-the-hood as a configurable parameter, allowing users to select a larger buffer as required to reduce the impact of cropping on extrapolated results. I've raised a separate issue for this one, as I think it would be a handy feature even beyond this specific issue: Allow cropping "buffer" as a customisable parameter #367

[tsutterley]

Hey @robbibt, I think option 2 will work really well. I'll start putting that together.