Project gaussians: adding numerical tests, resolving gradient bug

[vye16]

• added numerical tests for forward and backward passes against autograd on torch_impl in tests/test_project_gaussians.py
• fixed a gradient bug in projection. NOTE this changes the interface for the ProjectGaussians function, which now expects viewmat and projmat, rather than viewmat and the full matrix ofprojmat * viewmat as it did before.
• forward pass matches up to 1e-5 precision, backward pass matches up to 5e-4 precision.

[liruilong940607]

All looks great. I think the change on the interface of ProjectGaussians makes sense but we might need to consider backward compatibility. Is there a way to find out if the user is passing in the fullmat as before and raise a warning (preferable option)? Another option is to remain the API unchanged but compute the projmat from fullmat and viewmat.