move motion related classes from experimental to the normal locations

We currently have 2 sets of motion-related code:

• created by @ctsoumpas, used by PoissonLogLikelihoodWithLinearModelForMeanAndGatedProjDataWithMotion, see https://github.com/UCL/STIR/tree/master/src/spatial_transformation_buildblock
• a (mostly) older version created by @KrisThielemans, include files and definitions. This has files for rigid motion (and Polaris tracking), but also non-rigid. The latter functionality largely overlaps with @ctsoumpas's version. The main difference is
  • it defines the "transpose" (really adjoint) warp, and therefore doesn't need the inverse transformation to optimise the log-likelihood (and indeed does it theoretically better)
  • it also allows incorporating the Jacobian of the transformation

I checked quite a while ago, and at that point @ctoumpas' version was faster. I don't really know why, but @rijobro is putting in changes to avoid repeatedly warping images for every projection of RelatedViewgrams, see PR #239 , which might fix this.

Note that the 2nd version is currently not compiled by default (you have to create a extra_dirs.cmake in local for that).

The question now is how we merge these 2 versions. This issue is to discuss this further.

Compare the motion code under src/spatial_transformation_buildblock with the files in src/include/local/stir/motion and src/local/motion; the latter is only compiled through local/extra_dirs.cmake. Read PR #239 for the related warping changes and assess the adjoint and Jacobian functionality. Done requires an agreed merge plan for the overlapping implementations.