Fixed logic in invwarp that detects out-of-FOV voxels when --in > --ref

It turned out invwarp was very slow. This was due to a failure to properly detect voxels in the inverse warp field that was out of the FOV of the original --ref. That meant that there as a sharp edge between the "valid" and the "invalid" parts of the field in the inverse field. These edges were then detected as singularities, and "nudged" by NEWIMAGE::constrain_topology, which took a very long time. It now correctly detects the "out of FOV" voxels and use a mean dilation from the "valid" to the "invalid" parts of the field. This makes it much faster, and also a little more accurate.