Avoid splitting when not necessary as it may incur in performance hit. Indeed,
splitting means more (tinier) goals to verify, hence more call to provers.

Instead of Classifier.

Also update the codebase accordingly.

- Add support for neural network as array in language.ml
- Translate csv record as vnnlib constraints in dataset.ml
- Add simple dataset normalization functionalities in dataset.ml
- Add vnnlib printer (as smtlib printer with direct real support)
- Add vnnlib driver for PyRAT
- Rework onnx and nnet parsers to not fail on non-mandatory information