This PR adds new model / dataset features for upgrading our ensemble. This includes

Models

• Logistic Regression (although it does not perform very well)
• functioning LSTMs (better than LR, but not as good as Electra)
• freeze Electra weights (this has a negative impact on performance as well)

Datasets

• subset-based training (only 3_STAR / only 2_STAR molecules - note that training on a subset impacts performance when tested on a different subset))
• id / label filters - these come in handy when comparing datasets. I have used them to compare 3-STAR trained and full-ChEBI trained models (e.g., I took ChEBI50 with a label-filter from ChEBI50-3STAR to evaluate a 3-star model)
• PubChem batched dataset - this allows doing a PubChem dataset not as one file, but as batches. Training happens with a 1 batch per epoch scheme. I used this to pretrain a model on the whole PubChem dataset (118 million SMILES, 1 million per batch)