Make bkbit usable straight after pip install (no taxonomy pre-step)

[puja-trivedi]

bkbit was unusable after pip install: genome_annotation_translator loaded the NCBI taxonomy in the Gff3 class body, so a missing cache raised at import time and broke every subcommand, including --help. The cache was also 225MB of pretty-printed JSON written into site-packages, which is too large for PyPI and may be read-only or wiped on upgrade.

Taxonomy access now goes through bkbit.utils.ncbi_taxonomy_cache, which serves lookups from two layers and does no work at import time:

1. A subset bundled in the wheel (ncbi_taxonomy_data/taxonomy_subset.json.gz, 575KB, the 30,618 taxa that have a GenBank common name). Covers every BICAN organism, loads offline in milliseconds.
2. The full NCBI taxonomy, downloaded only on a subset miss and cached in a per-user cache dir (BKBIT_DATA_DIR to override, BKBIT_NO_DOWNLOAD to forbid the implicit download). Written compactly; legacy in-package caches are still read but never written to.

download-ncbi-taxonomy becomes an optional prefetch for air-gapped, CI, and container builds, and gained --data-dir. ncbi_taxonomy_data/build_subset.py regenerates the bundled subset for a newer taxonomy dump.

Also fixes two further import-time crashes on the same path: import pkg_resources fails on Python 3.12 with modern setuptools, and it was reached from cli.py via HMBA_annotation_translator (unused import, removed) and add_dunderMethods_genomeAnnotation (switched to pathlib).

pyproject.toml gains the runtime deps that were imported but undeclared (requests, tqdm, rdflib) plus platformdirs, declares package-data for the bundled subset, and requires Python >=3.9 for importlib.resources.files.

Verified by installing the built wheel (803KB) into a clean venv with an empty HOME: bkbit --help and bkbit gff2jsonld on a real Ensembl GFF3 both succeed with BKBIT_NO_DOWNLOAD=1 and no cache present.