Reduce S3 egress costs for bbop-sqlite database downloads

Problem

OAK downloads SQLite databases from https://s3.amazonaws.com/bbop-sqlite/{name}.db.gz. Every time a user's cache expires, the full file is re-downloaded even if nothing has changed on S3. For popular ontologies (GO, Uberon, HP, etc.) this generates significant egress costs.

Current state

• Default cache TTL increased from 1 week to 1 month (#879)
• No HTTP conditional requests (ETag / If-Modified-Since) — every refresh is a full download
• S3 URL is hardcoded in SqlImplementation.__post_init__

Proposed improvements (incremental)

Phase 1 — Decouple the download URL

• Introduce a configurable base URL (e.g. OAKLIB_SQLITE_BASE_URL env var or config) instead of hardcoding https://s3.amazonaws.com/bbop-sqlite/
• This lets us switch providers (CloudFront, R2, GCS, etc.) without a code release as pricing changes
• Could also support mirrors or institutional caches

Phase 2 — HTTP conditional requests

• Use If-Modified-Since / If-None-Match (ETag) headers when checking for updates
• S3 already returns Last-Modified and ETag headers, so a HEAD or conditional GET would return 304 Not Modified with near-zero egress when the file hasn't changed
• This means we could shorten the cache TTL again without cost penalty

Phase 3 — Delta/incremental updates (longer term)

• Explore whether we can avoid downloading full .db.gz files entirely
• Options: rsync-style deltas, SQLite diff/patch, or a lighter index file that indicates what changed

Related

• #879 — increased default cache TTL to 1 month

Start at SqlImplementation.post_init, where the S3 URL is hardcoded, and trace the existing cache and download path. First agree which proposed phase is in scope and how configuration or conditional requests should be tested; done should include the selected behavior without requiring the later delta-update exploration.