feat: serde + caching

[cmdoret]

Context

Loading and filtering the ontology when instantiating a TermMatcher object
is by far the slowest operation in (py)fuzon.

Summary

This PR adds efficient (de)serialization support to fuzon::TermMatcher using postcard, a low overhead binary serializer.

Changes

• Implement SerDe on TermMatcher
• Implement caching in CLI to prevent re-downloading the same ontologies over and over again.
  • cache keys are built from the path, last-modified timestamp and ETag(url) / size(file) of ontologies
  • if there are multiple ontologies, their data are sorted+concatenated before hashing
• Implement pyo3 bindings to TermMatcher::load/dump in pyfuzon
• Expose load/dump functionality in python API

Notes

Basic tests with the uberon ontology showed the following:

• The full ontology is 89MB
• Once filtered (by the TermMatcher constructor), we are down to ~2MB
• Creating a TermMatcher from scratch with locally stored UBERON took 2.765 (on NVMe).
• WIth postcard, deserializing a TermMatcher takes 65ms, the object dump takes 2.1MB

Current limitations

• There is currently no way to cleanup the cache (besides deleting the cache directory). Would it make sense to add it as a CLI flag?

[marftn]

LGTM :)