This fixes most of the cases, the only situation not addressed is when
the highpoly object(s) has non-uniform scale.
mul_transposed_mat3_m4_v3() should take care of non-uniform scales so
I'm a bit confused on why it doesn't work. The lowpoly object can have
any transformation, the only issue is if the highpoly object has
non-uniform scale.
Test file of the remaining issue:
https://developer.blender.org/file/info/PHID-FILE-tpw2xgddyzxtpg3e7xzs/
Reference reading:
http://www.unknownroad.com/rtfm/graphics/rt_normals.html