fix: enhance service health notebook with detailed topology descripti…

[shackmann]

…on and caching improvements

---

Note

Low Risk
Notebook-only documentation and diagnostic output; no runtime library or service behavior changes in this diff.

Overview
Expands service_health.ipynb so multi-deployment setups are easier to diagnose without duplicating health traffic.

The intro now documents how JOINTFM_DEPLOYMENT_IDS round-robin pools work: health() aggregates consensus metadata (notably minimum reachable max_sample_count for sample-batch splitting), and cache=True on the first fetch lets the new topology section reuse that probe.

A Deployment Topology section calls health_instances(cache=True) and prints selector, topology_label, parallel capacity (sum of caps) vs sample-batch cap (minimum), plus per-deployment availability, predict URL, device, image/checkpoint versions, and errors for unreachable peers.

The health fetch cell uses health(cache=True, refresh=True) and drops the redundant trailing cell output; the notebook kernelspec is renamed to Python (joint-client-python).

^{Reviewed by Cursor Bugbot for commit f64a8be. Bugbot is set up for automated code reviews on this repo. Configure here.}