neuralnetlib 3.3.7

• docs: remove useless comments
• fix(Transformer): too much things to tell
• feat: even more precise floating point for metrics and loss
• refactor: special tokens now passed via init for Transformer
• feat: enhance beam search and token prediction mechanisms
• docs: update readme
• fix(Transformer): vanishing gradient fix
• fix(Transformer): still on it (wip)
• fix(Transformer): another fix
• fix(Transformer): special token indices
• fix(Transformer): normalization IS the issue
• docs: update readme
• fix(Transformer): cross attention weights
• fix: LearningRateScheduler
• fix: LearningRateScheduler
• fix: normalization in data preparation
• fix: different vocab size for different tokenizations
• fix(PositionalEncoding): scaling
• fix(AddNorm): better normalization
• fix(TransformerEncoderLayer): huge improvements
• perf(SequenceCrossEntropy): add vectorization
• fix(Tokenizer+Transformer): tokenization alignement for special tokens
• fix(transformer): investigate and address gradient instability and explosion
• fix(sce): label smoothing
• refactor: gradient clipping
• fix(Transformer): gradient explosion
• fix(Transformer): tokens padding and max sequence
• test: tried with a better dataset
• fix(sce): y_pred treated as logits instead of probs
• fix(TransformerEncoderLayer): remove arbitrary scaling
• fix(Transformer): sce won't ignore sos and eos tokens
• fix: sce extending lossfunction
• fix(sce): softmax not necessary
• feat: add BLEU, ROUGE-L and ROUGE-N scores
• fix: validation data in fit method and shuffle in train_test_split
• docs: modifies example to use validation split and bleu score
• fix(PositionalEncoding): better positional scaling
• ci: bump version to 3.3.7