[OlivierBeq]

Summary

• Fix crashes in qsar()/pcm(): insufficient training data on non-year splits, and a KeyError from already-dropped columns during cluster split
• Fix DNN issues: checkpoint reuse, device kwarg collision in constructors, missing-target masking in multi-task loss, and mold2 sentinel NaNs in the advanced querying example
• Fix data loading: coerce/cast Year column to numeric/int32, raise a clear ImportError when torch/skorch are missing, add download_if_missing to PapyrusDataset.from_dataframe
• Suppress a harmless cupy CUDA path warning on FPSim2 import
• Add a protein descriptors example notebook and result counts to similarity/substructure search examples
• Document Windows CUDA wheel install for the dnn extra
• Ignore generated FPSubSim2 database files and notebook output artifacts

Test plan

• tests/test_modelling.py and tests/test_neuralnet.py cover the qsar/pcm and neural net fixes
• tests/test_oop.py / tests/test_oop_unit.py cover from_dataframe changes