add support for writing parquet files

[Hoeze]

This PR adds support for writing parquet files.
In particular, it adds two additional parquet writer implementations:

• ParquetFileBatchWriter: Writes to a single parquet file and depends on PyArrow.
• ParquetDirBatchWriter: Writes a directory of parquet files using pandas.to_parquet(). Supports both PyArrow and fastparquet. This implementation is predestined for multi-processing / asynchronous writing since no file-locking is necessary for it. Nevertheless, I'm not sure yet how Kipoi would take advantage of that.

My very long-term goal is that we directly provide schemas as Arrow schemas to be able to pass them around cross-language without any kind of conversion/serialization:
https://arrow.apache.org/docs/python/api/datatypes.html
This PR is the first step in that direction.
Future work:

• Explicit Arrow schema conversion
• Use Tensor type for arrays (GH-15483: [C++] Add a Fixed Shape Tensor canonical ExtensionType apache/arrow#8510)
• Implement struct/list types in Pandas (Implement nested types: Add StructDtype and StructArray pandas-dev/pandas#45745), e.g. for GenomicRange

[haimasree]

Thanks a lot! Very cool! Try with adding pyarrow. Once the tests pass I will merge to master

[haimasree]

So CI is telling me that pyarrow is not installed. I dont see it in any dev-requirements-py*. yml either. Perhaps try adding pyarrow in kipoi setup.py?

[haimasree]

The exact error is ModuleNotFoundError: No module named 'pyarrow'

[Hoeze]

Thanks! Yes, I have to add it as an optional dependency 👍

[haimasree]

I will merge this and remove the coverall for now. I will check tomorrow why the REPO token is not working anymore.