Similar to the problem in LevMarq, arg.inl was being used instead of
best.inl. This opened us up to a potential segfault.

Listed accesses for more functions.

Spaced methods & functions more consistently, and started documenting
which members does each method access directly or through its callers
within RHO_HEST_REFC.

- Deleted "RefC" from names of external-interface functions.
- Renamed rhorefc.[cpp|hpp] to rho.[cpp|hpp]
- Introduced RHO_HEST base class, from which RHO_HEST_REFC inherits.
- rhoInit() currently only returns a Ptr<RHO_HEST_REFC>, but in the
future it will be allowed to return pointers to other derived classes,
depending on the values returned by cv::checkHardwareSupport().

The call to rand()/random() is now a call to (unsigned)cv::theRNG().

Cholesky decomposition is stable; It is not necessary to carry it out
internally at double precision if the result will be truncated to single
precision when stored.

- Switched to the extremely fast, while simple and high-quality,
xorshift128+ PRNG algorithm by Sebastiano Vigna in "Further scramblings
of Marsaglia's xorshift generators. CoRR, abs/1402.6246, 2014" (2^128-1
period, passes BigCrush tests). Performance improved by 10% over
random().
- Added an API to allow seeding with a specified seed, rather than using
rand() or random(). This allows deterministic, reproducible results in
tests using our algorithm (although findHomography() does not yet
support passing an entropy source on its own end).

Fix the issue given NULL inlMask

Fix the issue given NULL inlMask

Fix the issue given NULL inlMask

Fix the issue given NULL inlMask

Fix the issue given NULL inlMask

The code has been refactored in response to feedback on Pull Request

Also, outputZeroH() now also zeroes the inlier set, much like
outputModel().

Previously, the output mask of inliers could remain completely
uninitialized. This fix is the first part of a solution.

The inverted JtJ does not multiply itself, but rather Jte. Correct this
in the comment.

Added explicit casts to silence warnings in fundam.cpp and rhorefc.cpp.

Implemented a damping-parameter choice strategy similar to that
described in http://www2.imm.dtu.dk/documents/ftp/tr99/tr05_99.pdf.
Removed a few debug statements.
Chosen a new starting lambda value, 0.01.
We now actually output the mask of inliers.

Replaced the complex rules OpenCV uses to select lambda with a naive but
fast heuristic. It's imperfect but produces good results. It is still
subject to the same problem as OpenCV - namely, on occasion LevMarq will
make a poor result even worse.

Deleted multiplications by zero and consequent additions of zero terms.

LevMarq now doesn't outright fail, but doesn't seem to improve things
much if at all.

Refactoring of Cholesky decomposition.
Fix for memory corruption bug.
LevMarq as a whole still non-functional.

Implements the RHO algorithm as presented in

Paper: Bilaniuk, Olexa, Hamid Bazargani, and Robert Laganiere. "Fast
Target Recognition on Mobile Devices: Revisiting Gaussian Elimination
for the Estimation of Planar Homographies." In Computer Vision and
Pattern Recognition Workshops (CVPRW), 2014 IEEE Conference on, pp.
119-125. IEEE, 2014.

- Complete, heavily documented reference C implementation, as well as
temporarily disabled dirty SSE2 port.
- Enabled tests for RHO in test_homography; Currently these fail
presumably due to too-stringent accuracy requirements.
- Refinement and final refinement are not yet functional; Do not pass
their corresponding flags to RHO.