Non-exhaustive list of possible optimizations and refactors

[MatthieuHernandez]

• Factorize back(float error) and train() between SimpleNeuron and RecurrentNeuron
• Maybe rename computeBackOutput and computeTrain
• What about this->errors in operator==?
• Default operator==
• Add ResetLearningVars
• Remove learning vars from Archive
• Add a real batchSize()
• Make literals operator consteval
• previousOutput not used in GRU
• Tensor class can be made with private default constructors and friend to boost or something like that.