Agent-Friendly Client + Metadata Service Foundation

Metaflow's metadata service returns unbounded JSON for every list endpoint. A single GET /runs on a production deployment returns megabytes of data the caller immediately discards. For AI agents, this is worse: a failure-diagnosis workflow generates 2,021 HTTP calls to check 1,000 tasks. This proposal adds cursor-based pagination and server-side tag filtering to the metadata service (prototyped in PRs #9 and #10, 65 tests passing), integrates pagination into the Python client SDK, and builds agent-friendly query utilities that, across five common agent workflows, collapse 11,828 API calls into 16. Server-side work is already prototyped. GSoC starts at hardening, not greenfield.