Fixes #82642

The issue

It's very common to get warnings when baking LightmapGI and it's easy to get tens of thousands of them in a single scene.

If the scene is smaller than about 100x100 meters, these warnings are often harmless and don't have a noticeable effect on the indirect lighting of dynamic objects. But for larger scenes, the effect is very much noticeable.

The warnings and the bad accuracy have three causes:

1. A float precision problem in Plane.
2. A common and valid scenario where a partitioning plane can't be found for the BSP tree.
3. Delaunay3D can create invalid triangulation with overlapping tetrahedra in scenes covering a large area.

I made #90701 a separate PR for the last issue since Delaunay3D can be accessed directly from GDScript so that change isn't strictly isolated to LightmapGI.

Increased tolerance for Plane::has_point()

When creating a Plane from a triangle that spans hundreds of meters, Plane::has_point() can give false negatives when applied to the second and third point of the triangle with the default epsilon. And is_point_over() can return either true of false in these cases.

This wreaks havoc on BSP partitioning.

I've done tests on MRPs and scenes with randomly placed probes over 1000x1000x100m. The precision errors are always a power of two and the largest one I've been able to produce is 1/(1<<13) so I'm using that for tolerance when calling has_point(). It is 12x the current tolerance of 1e-5 but still small enough that it shouldn't matter. False positives with this magnitude should not have any effect.

A better partitioning fallback

Probe tetrahedra are recursively sorted into a BSP tree. All the faces of the tetrahedra are candidate planes to be used for partitioning. As long as the remaining tetrahedra in a subtree have at least two that share an edge or a face, this method will always find a plane that cleanly separates at least one into each new subtree.

But the final two (or very rarely three) tetrahedra in a branch are often either disjoint or share only a single vertex. In this scenario, it is no longer guaranteed that one of the faces is also a separating plane. This is common cause of warnings.

With this knowledge, we can handle this specific case explicitly in a way that always succeeds (as long as the triangulation from Delaunay3D is valid).

Bonus: Smaller BSP tree

When the chosen BSP separating plane intersects tetrahedra, they are not split but instead moved into both the subtrees.

The current scoring system to find the best plane only rewards balance. If a plane has one tetrahedron over it, one under and 20 intersecting then it gets a perfect score because 1/1 is perfect balance. This makes the BSP tree much larger than it needs to be.

Reduz actually had a better solution but he commented it out. I took it and I made it more extreme. The scoring now penalizes planes for the number of tetrahedra they intersect.

"Worst" case, the tree is now 30% smaller. Best case, 97% smaller.

I can go into details about why I'm not worried that this will create unbalanced trees, if someone is interested. I've also verified that the max tree depth on all my test scenes is less than before. I expect that lookup times at runtime in general should "always" be improved. And when there are many probe nodes, bake time is also noticeably reduced.