If raster contains a single NaN Linear() works but Cubic() returns all NaNs

I'm seeing issues when using Cubic interpoaltion of data that contains NaN values. Here's what I see:

```julia
using Interpolations
A = rand(100,100);
x = 1:100;
y = 1:100;
interp = Cubic();

itp = Interpolations.extrapolate(Interpolations.scale(Interpolations.interpolate(A, BSpline(interp)), x, y), NaN);
```

itp has no NaN values
```julia
julia> sum(isnan.(itp.itp))
0
```

add a NaN value
```
A[1,1] = NaN
itp = Interpolations.extrapolate(Interpolations.scale(Interpolations.interpolate(A, BSpline(interp)), x, y), NaN);
```

now all itp values are NaNs
```julia
julia> sum(isnan.(itp.itp))
10000
```

change interp to Linear
```julia
interp = Linear();
itp = Interpolations.extrapolate(Interpolations.scale(Interpolations.interpolate(A, BSpline(interp)), x, y), NaN);
```

now itp contains only a single NaN
```julia
julia> sum(isnan.(itp.itp))
1
```