Trainer class

[Ithanil]

This PR contains a NN training system, the updated smart beta generator and a cleanup of directory structure and scripts. Also contains 3 new unittests for the new code (including your smart beta unittest) and an example for the trainer. In total we have the following new classes/files: NNTrainingData, NNTrainingConfig, NNTrainer, NNTrainerGSL, GSLFitStruct, SmartBetaGenerator.

Besides the directory/script changes, everything should be well explained by respective doc sections, slack communication and similar.

Directory/script changes:

• The src folder is now nested, with one more level of deepness: e.g. the former src/InputLayer.hpp is now src/layer/InputLayer.hpp
• Compiler will now work in bin/, so .o files will go there
• To keep the easy including without folder prefix, we now have a flat include directory with symlinks to all header files (we got a mini script to update the symlinks after adding/removing a header file)
• All scripts went into the new script folder
• Except the actually new build.sh (still root), which now takes an argument that can be "opt", "debug", "profiling" or "coverage". It will call the appropriate build script in script/. Without argument it will build opt.
• New common compile script in examples directory. The individual example run scripts source it.

EDIT: Obviously the whole build/include management is becoming increasingly complex. I think it is time to make the transition to an autotools-based build (i.e. ./configure && make). But this will be added in a later PR.

EDIT2: In the meantime, further significant changes have been introduced with this PR: 1) Even faster ray implementation

I did another overhaul of the feeder implementation (especially the isBetaUsed problem) and think we have now reached a really clean and reasonable final version (and another 30% performance boost).

2) Betas vs. Variational Parameters

• Now there is a clear distinction between betas and variational parameters. The set of "betas" remains to mean all weight factors used by all feeders. However, the set of variational parameters means the set of overall parameters (which may include a subset of betas) chosen to be varied by an (external) optimizer. You need to make a parameter a variational parameter, if you want to be able to access it via any "getVariationalParameter()" methods and similar or if you want to extract a variational derivative for it (i.e. vd1, cd1, cd2).
• Currently the only way to do this is via FFNN's "assignVariationalParameters(const int &starting_layer_index)", where starting_layer_index denotes the layer from which we start to assign betas to variational parameters (i.e. all betas from this layer and all layer layers after it will be added).
• If you add variational/cross substrates, only substrates for variational parameters will be added (not necessarily all betas/parameters). This made the LastHiddenlayer addSubstrate methods obsolete and they were removed.
• The variational parameter status will be (inherently) correctly restored by constructing from file / copy constructor.

[coveralls]

Pull Request Test Coverage Report for Build 98

• 706 of 788 (89.59%) changed or added relevant lines in 25 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage increased (+83.6%) to 83.597%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/actf/ReLUActivationFunction.hpp	0	2	0.0%
src/actf/SELUActivationFunction.hpp	0	2	0.0%
src/actf/TanSigmoidActivationFunction.hpp	0	2	0.0%
src/feeder/SmartBetaGenerator.cpp	98	100	98.0%
src/actf/SineActivationFunction.hpp	0	4	0.0%
src/io/PrintUtilities.cpp	0	11	0.0%
src/trainer/NNTrainer.cpp	93	110	84.55%
src/network/FeedForwardNeuralNetwork.cpp	67	88	76.14%
src/trainer/NNTrainerGSL.cpp	232	253	91.7%
<!--	Total:	706	788	89.59%	-->

Totals
Change from base Build 96:	83.6%
Covered Lines:	1641
Relevant Lines:	1963

---

💛 - Coveralls