Add fp16 support for sample_farthest_points, ball_query and knn

[eamonn-zh]

add fp16 CUDA supports for sample_farthest_points, ball_query functions via AT_DISPATCH_FLOATING_TYPES_AND_HALF and knn

[bottler]

(There's no need to keep merging from main. There are no conflicts afaict.)

This could be a useful addition - do you have an application?

My concern here is that it makes builds bigger and slower, which is the opposite of where we really want pytorch3d to be.
A big question: Could these operations be reimplemented in python and be very efficient with torch compile? Or could they be written in openai triton? Or, if the answer is no, could they be compiled on-demand with torch.utils.cpp_extension.load?