Also removed trailing space from vertex/colour data

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.

Renamed fit_shape_model to fit_shape, since it's fitting the shape in general (including blendshapes), and not only the PCA model.

Added a BUILD_UTILS CMake option, defaulting to OFF.

- Added expression fitting
- Added Surrey Face Model section
- Slightly restructured build and sample section
- Replaced ICIP with the VISAPP publication
- Replaced license badge with static one

Related to last commit: Added serialisation methods for boost::optional. Copied the file from https://github.com/USCiLab/cereal/pull/122, hopefully that pull-request will be accepted.
This commit adds the file to CMakeLists.

I don't think we need two representations for it, one struct with a type is easier. At the moment, we only use an orthographic camera anyway.

- As before, it should contain all the parameters necessary to render an image using OpenGL, and conform to the OpenGL conventions
- Updated all calling sites
- Added serialisation
- Added a convenience method to save the parameters to a json file

Probably needed because OpenCV needs it, and we depend on OpenCV.

The previous code is non-standard conformant, it was assigning a temporary to a non-const reference.
VS2015 accepts it most likely because of a non-standard extension.

* The inner mouth corners are not mapped since I think they don't realy exist in the 3D model.
* Some points are not defined on the reference and only present in the 3448 model and above.
* A few of the selected 3D vertices could be improved a bit, I have some notes on it, if someone is interested, contact me.

If not given, the mean will be used, which is the right thing in that case.

This is used for example when we fit blendshapes and want to fit the PCA shape given current blendshape parameters.

See GitHub issue #4 for details.

This runs 10-30% faster (and is potentially more stable).

Based on the linear landmark fitting.
The regularisation is set conservatively, in the experiments so far even a zero value yielded good results.