fix(server): refresh stale point lookups across stores

[beinan]

Summary

• move explicit dataset-version pin tracking into shared StorageBase, so refresh behavior is consistent for context, rollout, generic, and datagen stores
• keep point-read hits on the existing shared-lock fast path; on a miss, acquire the write lock, checkout_latest(), and retry once
• cover context id/external-id/payload reads, rollout record/blob reads, generic row reads, and datagen fold/event/status/failure/blob reads
• preserve explicit context and rollout time-travel checkouts instead of auto-advancing pinned handles
• add deterministic multi-handle regressions where an external writer merges a row, drains the WAL, and leaves the cached reader on an older base version

Testing

• ./.codex/skills/ci-pr-helper/scripts/run_ci_checks.sh
• Rust core: 217 passed, 3 ignored
• Rust server: 67 passed
• WAL integration: 6 passed
• Python: 213 passed, 6 skipped, 3 xfailed
• cargo fmt/clippy, ruff format/check, and pyright passed