Adjusted RenderingParameters constructor accordingly.

This is useful to get correct angle estimates when, as in OpenCV, the viewport is upside-down.
It could maybe also be solved by negating the 2nd and 3rd row of the rotation matrix in RenderingParameters, but I couldn't get the projection right. This is probably nicer.

It still needs a more proper cleanup.

RenderingParameters gets more and more chaotic, needs a major overhaul.

Based on the affine camera estimate, and then doing SVD to obtain the closest orthonormal rotation matrix.

This is mainly to accommodate fit-model-ceres, where I've used double for all parameters

At least gcc-5.4 on Ubuntu 16.04 needs it. It shouldn't hurt on older versions.

Changed all affected functions and classes accordingly: Mainly the renderer, affine-renderer, and affine texture extraction.

This results in a ~35% speed-up of rendering meshes.

It requires some not-so-beautiful conversions in render_affine and the current texture extraction, but they will be superseded by the new renderer soon anyway.

Fits camera, shape, blendshapes and colour coefficients with a cost function containing a landmark constraint, priors, and minimising the error between the input image's RGB values and the model estimate (rendered at each vertex).

Note: The image cost function cannot be used with the model on GitHub, as it doesn't include the colour PCA.

Not sure what's the best place to put them, let's put them here for now.

* Fits the pose (camera) to landmarks using the newly added Ceres cost functions
* Added find_package(Ceres) to CMakeLists.txt

He updated it quite a few months ago and sent the file to me - now finally committed.

* The changes were necessary to fix issues with Ceres's Jet type in glm vector and matrix types
* Added glm-git-modified-readme.txt to explain these changes

* Added const and non-const overloads. The non-const is needed because of the RNG in PcaModel.
* Renamed a few variables for consistency

This is needed by the json-to-cereal-binary app to read the JSON generated from the Matlab BFM converter script.

Otherwise, it looked like JSONlab will not write the output file, probably because it's >1GB. See GitHub issue #29.