This commit should solve two bugs:
1. Do not get stuck in infinite loops on certain shapes
2. Offset the entire outline by 0.25px outside along the normal to avoid
duplicate points in 1px wide areas (duplicate points can not be accepted
by the meshing algorithm)

* GEGL now depends on poly2tri-c (git commit 566e50 or later)
* The old version of poly2tri-c was removed from GEGL's repo

The change was only applied to the gegl:seamless-clone operation, and it
should later be applied to the gegl:seamless-clone-render operation.

Instead of using the environment variable p2t_refine_steps, the limit is
now a parameter to all the functions, and is also a parameter of the
gegl operation itself. It won't crash now if the environment variable is
missing (that was the state previously).

Since we can't compute the color difference of an outline point if it's
outside the background bounds, we will simply ignore it in the weighted
average. This removes the artifacts that appeared from such cases.

Note that this doesn't solve the problem where all selected sample
points are outside the background, but the algorithm isn't really meant
for handling such cases so this is the best we can do.

In the seamless-clone-prepare op, warn if the destination pointer for
the result is NULL. Also, actually store the pointer to the result at
the given pointer to a pointer.

In the seamless-clone-render op, rename the offset properties to "x"
and "y" instead of "xoff" and "yoff".