Comparison study of new vs. legacy model implementations

[asogaard]

Have added:

• separate zenith and azimuth reconstruction tasks to better reflect their different natures,
• an example script for training new and legacy models; closes Train 2D and 3D vonmisesfisher on zenith and azimuth( 4 models in total) #55
• an example notebook for visualising a few results; closes Compare 3D and 2D vonmisesfisher azimuthal prediction distributions. #54, closes Compare zenith distribution, sliced in kappa, for exact von Mises-Fisher loss (m=2) vs. approximation with m=3 for single angle #44
  The punchline, as far as I can tell, is the performance is about the same, but the exact von Mises-Fisher loss yields better (i.e. more-unit-variance-y) pulls and are easier to motivate in a paper.

@RasmusOrsoe, let me know if you find similar results using your own plotting scripts.

[RasmusOrsoe]

Thanks! Will get back to you