feat(03-06): implement literature evidence models, PubMed fetch, and scoring

- Create LiteratureRecord pydantic model with context-specific counts
- Implement PubMed query via Biopython Entrez with rate limiting (3/sec default, 10/sec with API key)
- Define SEARCH_CONTEXTS for cilia, sensory, cytoskeleton, cell_polarity queries
- Implement evidence tier classification: direct_experimental > functional_mention > hts_hit > incidental > none
- Implement quality-weighted scoring with bias mitigation via log2(total_pubmed_count) normalization
- Add biopython>=1.84 dependency to pyproject.toml
- Support checkpoint-restart for long-running PubMed queries (estimated 3-11 hours for 20K genes)

pyproject.toml

@@ -35,6 +35,7 @@ dependencies = [
     "pyyaml>=6.0",
     "httpx>=0.28",
     "structlog>=25.0",
+    "biopython>=1.84",
 ]
 
 [project.optional-dependencies]
@@ -42,6 +43,9 @@ dev = [
     "pytest>=7.4.0",
     "pytest-cov>=4.1.0",
 ]
+expression = [
+    "cellxgene-census>=1.19",
+]
 
 [project.scripts]
 usher-pipeline = "usher_pipeline.cli.main:cli"