Support for large Field-of-View cameras in the robust estimators/refinement

[vlarsson]

While most minimal solvers gracefully handle large FoV (>180 deg) the current robust estimators make heavy use of normalizing to a 2D image plane which works poorly for these cameras. Additionally, the camera models are modeled as a non-linear mapping applied to the pinhole projection which also breaks in this case.

Adding support for these models would require many things to happen

• Support for projection/unprojection which goes from 2D image to 3D bearing vector and vice versa
• Update RANSAC estimators to use camera model instead of normalized coordinates which is the default now
• Rewrite Jacobians in jacobian_impl.h to support this as well

Since this requires quite significant changes to many parts of the code, I think it's probably best to think on this for a while and come up with a concrete plan before starting to hack away...