Add new TRS functionality. (#580)

Summary:
Pull Request resolved: #580

We are finding that many users of the PyMomentum library prefer the use of rotation matrices to quaternions because rotation matrices tend to function better in ML learning algorithms.  However the existing skel_state matrix functionality converts to 4x4 matrices which is challenging to work with because scale has been baked into the matrix and needs to be removed.  Therefore we decided it likely make sense to have support for a TRS format that stores separate translations, 3x3 rotations, and uniform scales.  This format can be easily converted to-and-from skel_states, and is generally safer to work with than 4x4 matrices (for example inversion just involves a transpose).

Reviewed By: jeongseok-meta

Differential Revision: D82840072

fbshipit-source-id: 22b8d2d4294231a7af5fccbb72c9f0243d54a479

Author: cdtwigg

pymomentum/cmake/build_variables.bzl

@@ -135,6 +135,10 @@ skel_state_sources = [
     "skel_state.py",
 ]
 
+trs_sources = [
+    "trs.py",
+]
+
 marker_tracking_public_headers = [
 ]