fix for bilinearInt NaN handling

[terrry-sh]

Hey guys,
I've picked up an error in the handling of NaNs in the bilinearInt method for the extract method.
See the example below:

r <- terra::rast(nrow = 2, ncol = 2, nlyrs = 1, xmin = -100, xmax = 100, ymin = -100, ymax = 100)
terra::values(r) <- c(100, 100, NaN, 100)
print(terra::extract(r, data.frame(X = 45.0, Y = 45.0), method = "bilinear")$lyr.1)
# ought to be 100 but the value will be 55

Here the expected bilinear interpolation between points all with value 100 should always stay at 100 but we dip down to 55, 145. A quick peek at the source code shows that the handling isn't linearly adjusting when a NaN is found and simply using the value 0.5 which can lead to problems say when (y - y2) happens to be very small or very large; see below

terra/src/extract.cpp

Lines 168 to 183 in 1820db4

	if (intx & inty) {
	double d = dx * dy;
	if (!(n1 || n2)) {
	w11 = v11 * ((x2 - x) * (y - y2)) / d;
	w12 = v12 * ((x - x1) * (y - y2)) / d;
	} else {
	w11 = n1 ? 0.0 : 0.5 * v11;
	w12 = n2 ? 0.0 : 0.5 * v12;
	}
	if (!(n3 || n4)) {
	w21 = v21 * ((x2 - x) * (y1 - y)) / d;
	w22 = v22 * ((x - x1) * (y1 - y)) / d;
	} else {
	w21 = n3 ? 0.0 : 0.5 * v21;
	w22 = n4 ? 0.0 : 0.5 * v22;
	}

I've implemented in the first commit a few tests which fail on the current version of extract with method = "bilinear" and then in the second commit I've added the linear/bilinear factor back in to fix this issue. All tests pass and personally I've been using this algorithm on rasters at work with good results.

[rhijmans]

Thank you very much. I have merged it (but have not had time to give it a good look)

[terrry-sh]

Thanks!