Mot

====================

Added
-----
- Adds CL context cache to fix issue 5.
- Adds singularity boolean matrix to the output of the Hessian to covariance matrix.

====================
- Simplified the CL context generation in the hope it fixes issue 5.

====================

Changed
-------
- Changed the default load balancing batch size.

===================

Added
-----
- Numerical Hessian now with OpenCL support
- Adds method to get the initial parameters of a model.
- Adds initial lower and upper bound support to the numerical Hessian method.
- Adds a method to the sampling statistics to compute the distance to the mean.
- Adds InputDataParameter as superclass of ProtocolParameter and StaticMapParameter.
- Adds support for restrict keyword in CL functions.

Changed
-------
- Updates to the numerical Hessian calculation, translated more functions to OpenCL.
- Updated the buffer allocation in some methods to the new way of doing it.
- Updates to the numerical Hessian calculation, small improvement in local workgroup reductions.
- Changed the interface of the input data object to get the value for a parameter using a method call.

Other
-----
- Sets the default step size to 0.1 for the numerical differentiation, small updates to the numerical Hessian computation.
- Most of the numerical Hessian computations are now in OpenCL. Only thing remaining is median outlier removal.
- Made the KernelInputDataManager smarter such that it can detect duplicate buffers and only load those once. Furthermore, KernelInputScalars are now inlined in the kernel call.
- Made the method wrapping in the wrapped model easier.
- Lets the random restart use the model objective function instead of the L2 error. Furthermore, removed residual calculations in favor of objective function calculating.
- Renamed EvaluationModels to LikelihoodFunctions, which covers the usage better.
- Removed the GPU accelerated truncated gaussian fit since it was not doing the right thing. Added a MLE based truncated normal statistic calculator.
- In MCMC, changed the order of processing such that the starting point is stored as the first sample.

===================
- Small fix to the work group size, this will fix a INVALID_WORK_GROUP_SIZE issue with the procedure runner.

===================

Added
-----
- Adds a GPU based truncated gaussian fit.
- Adds a GPU based univariate ESS algorithm.

Changed
-------
- Updates to the model function priors.
- Updates to the KernelInputDataManager.
- Changed the sample statistic to use the CPU again for the easy statistics, for large samples this is faster than using the GPU.
- Updates to the function evaluator, made the input argument r/w by default and allows for void output functions.

Other
-----
- Prepared new release.
- Refactored the residual calculator, small performance update in MCMC.
- Removed two old mapping routines, the objective calculators.
- Project renaming.
- Work on the log likelihood calculator.
- Simplified some sampling post processing after changes in MOT.
- Removed the GPU multivariate ESS again, it was only marginally faster.
- Small speed update to the GPU univariate ESS method.
- More work on the procedure evaluator. Moved more data management tasks to the kernel input data manager.
- Renamed CLHeader to CLPrototype, covers the usage better.