Add metadata-driven Croissant parsers for HGNC and BindingDB

[cbizon]

Summary

• add a metadata-driven parser framework that combines Croissant metadata with ORION parser specs
• integrate HGNC as the default metadata-driven parser and add BindingDB as a metadata-driven Croissant loader
• fix live BindingDB parsing issues so legacy and metadata-driven parsers match on the 202603 archive

Validation

• uv run pytest tests/test_metadata_driven_parser.py
• uv run pytest tests/test_metadata_driven_parser.py tests/test_source_metadata.py
• ran legacy and metadata-driven HGNC on the live HGNC dataset; both produced 29,405 nodes and 34,075 edges
• ran legacy and metadata-driven BindingDB on the same BindingDB_All_202603_tsv.zip archive; both produced 1,200,861 nodes and 1,892,606 edges with identical node and edge SHA-256 hashes

Notes

• BindingDB comparison report: /tmp/orion_parser_runs_20260402/bindingdb_compare_202603/report.json
• HGNC run artifacts: /tmp/orion_parser_runs_20260402/hgnc_shared

[cbizon]

Check it out @EvanDietzMorris

I think that the current metadata for going data->graph highlights the concern that the metadata will need to be just as complicated as code. So not 100% sure it's a win, but maybe that can be fixed.