Before this commit it appeared that STL viewers could handle the output. However,
rendering the models in OpenGL revealed that the surface normals were all messed up. I
believe most STL viewers just don't care about surface normals, because I suppose they
don't really matter for 3d slicing.
The problem ended up being really a complex issue around the winding order of triangles in
sub-parts, and how this is affected by the "inversion" state. The winding order of
triangles in ldraw is a complex combination of BFC meta commands and transformation
matricies that "mirror" sub-parts. I tried to leave extensive comments in the code rather
than go into it in the commit message.
Also added a json output for rendering as OpenGL buffers. This is to help me debug issues
with the surface normals.
While working on this I also realized that the output of parsing a sub-part can be cached,
as long at the "invert" flag is part of the cache key. This could make a huge difference
in performance, because for example a "stud" gets generated many many times, and in this
version of the code that results in a lot of repeated processing.
I was concatenating paths together using the unix slash `/`. This change used
`File::Spec::catpath` instead, however I haven't been able to test it on Windows, and
there may be other path issues. Also note: without debugging on the user doesn't see when
files are not found (eg: due to path issues), so it silently creates an empty STL file.
