complete train/test split routine

[kjappelbaum]

Updates:

• we cannot read all datasets in chunked fashion. Luckily, those we can fit into memory. Therefore, we use a pythonic try/except to not chunk those datasets (this would probably also have caused issues with the chunked CSV via pandas setup)
• some datasets did not produce data_clean.csv when running transform.py
• some datasets did not have column names matching with what meta.yaml promised
• there were some dask issues with assigning partitions
• added script for checking for data leakage in case of the SMILES splits
• for some datasets (also not the largest ones) dask's type inference fails. We again try/except it pythonically

Also datasets that are more difficult to parse now do something meaningful

zinc has been causing issues for MicPie, also works

the LocalCluster from one of the commits here is not strictly needed, but it helps with debugging.

couldn't test on HPC as I still get Quota issues after deleting many files.

need to verify that we have all dependencies: openpyxl, pymatgen, givemeconformer, dask

[kjappelbaum]

remaining ran through

[kjappelbaum]

Still waiting for rdkit in the SMILES split, but otherwise this seems to run

[kjappelbaum]

one we need to check a bit more carefully is odd_one_out.

[kjappelbaum]

Since I need to redownload, I'll call it a day now while it continues running

[kjappelbaum]

odd one out also ran successfully, Iupac names is another large and slow dataset