One folder per dataset

Currently, we're grouping files per file type: (checksums/, fake_data/, datasets/).
We should try to move toward a more modular approach where datasets are self-contained directories:

dataset_type0/   # image, text, etc
    dataset0/
        dataset0.py
        dataset0_test.py
        checksums.tsv
        dummy_data/
    dataset1/
        dataset1.py
        ...

TODOs:

1. We should make it easy for external users to follow this model:
   Checksums path, fake data,... should be automatically detected without the user having to do the additional setup (tfds.core.add_checksums_dir, FAKE_DATA_DIR,...)

2. Update scripts like create_new_dataset, document_dataset, etc.

3. Update dataset testing methods

4. Update all files depending on checksums, fake_dataset locations etc

Start by tracing the current checksums/, fake_data/, and datasets/ layout, then inspect the create_new_dataset and document_dataset scripts and dataset testing methods. Map every dependency on checksum and fake-data locations before changing the organization. Done means datasets are self-contained, paths are detected automatically, scripts and tests support the new model, and all dependent files are updated.