Fix/prediction upload

[j279li]

This pull request introduces improvements to how Zarr object codecs and chunking are handled for prediction outputs, as well as some minor serialization and documentation fixes. The main focus is on making object codec support explicit and robust, especially for custom codecs, and ensuring correct serialization behavior for non-JSON-serializable fields.

Enhancements to Zarr object codec and chunking support:

• Added a utility function detect_object_codec_and_chunking in polaris/utils/zarr/codecs.py to determine the correct object codec, filter list, and chunking compatibility from template filters.
• Made codec chunking support explicit by adding a supports_chunking attribute to RDKitMolCodec (set to True) and AtomArrayCodec (set to False).
• Improved handling of None values in AtomArrayCodec.encode to explicitly set missing values in packed arrays.

Serialization and model improvements:

• Updated BenchmarkPredictionsV2 to exclude the non-serializable dataset_zarr_root attribute from JSON serialization using a Pydantic Field directive, and modified __repr__ to exclude the predictions field as they are also stored in zarr.

Minor documentation cleanup:

• Removed an outdated docstring reference to splits in BenchmarkV2Specification.

[cwognum]

In line with my comments in the Hub PR:

Rather than having these workarounds, I think it makes more sense to ditch the custom codecs. We can use default codecs (i.e. MsgPack for AtomArrays and VLenBytes for RDKit mols), to ensure it remains a valid default Zarr archive that any machine can open, and then convert from these formats to the objects we need internally.

This would be a bigger change than you've signed up for, though, as it also requires non-trivial changes to the dataset class. Let's do the following:

1. For datasets, we keep using the custom codecs
2. For predictions, we use default Zarr codecs and add conversion code on the client and service side.

Does that make sense?

[cwognum]

Hi @j279li , left some comments. Most are pretty minor! Take a look and let me know what you think.