Speedup dataset lazy loading

[wconnell]

1. MoleculeDataset(lazy=True) converts a smile to an rdkit Mol object and overwrites it with None. Loading large datasets like ZINC2m is severely slowed down by this, so avoid the unnecessary conversion to a Mol when lazy=True
2. Check download/extract logic with ZINC2m dataset to avoid redundancy and conflicts.

[KiddoZhu]

Thanks for your contribution! That's a really important step for large datasets.

For the check of file in Zinc2m, I would prefer keeping extracting the zip file anyway. This is because we don't check the md5 of extracted file and it is possible that the extracted file might be broken (e.g. due to Ctrl+C break).

[KiddoZhu]

Manually rebased and merged