Reproject PixelGridDefn object using dense bounding box edges (not just corners)

[neilflood]

Currently, footprint intersections etc., are calculated using reprojected corners of pixel grid bounding boxes. For some combinations of projection, on sufficiently large areas, the lack of the curved edges can result in clipping of the reprojected bounding box in ways which can lose data.

This PR implements reprojection of these bounding boxes using a fairly dense sampling of points along the edges, allowing the curve to be represented. This largely eliminates the problem, at least up to the level of the much shorter straight lines between sample points.

The following images demonstrate the problem and solution. The first circle is in lat/long coordinates. It is centred at (133E, 12.5S), i.e. in northern Australia. The radius is 2.5 degrees. The image is 42 degrees wide (roughly the width of Australia), which gives a long enough edge for the curvature to be significant.

The second is output from RIOS, without this PR. The working grid was in Australian Albers projection (3577), which curves the top edge to the north. Since the bounding box of the working grid does not allow for this curvature, the top of the circle is clipped off.

The third output is the same, but using the new implementation. As desired, the circle is not clipped.

Finally, this is the same image reprojected by gdalwarp, showing a very good match.

[neilflood]

This addresses some of the concerns raised in #204

[neilflood]

Thanks @gillins. I am just putting together a rigorous test of this code to include in the test suite.