Add always-on build telemetry delivery

[anth-volk]

Summary

• start a local telemetry emitter service automatically for supported US and UK dataset builds
• keep network access, Hugging Face authentication, retries, process sampling, and the bounded SQLite queue outside the build process
• use HF_TOKEN, HUGGINGFACE_TOKEN, or the cached Hugging Face login automatically; no telemetry configuration is required for PolicyEngine members
• exchange only the Hugging Face credential for one short-lived collector session credential, then register runs and upload events with that credential
• use one fixed production collector origin with no production destination override; retain only an explicit loopback option for development
• permanently keep a run local when no Hugging Face credential exists or membership is rejected, including across process restarts, while retaining transient failures for retry
• report stage transitions, graph-node and batch progress, calibration progress, resource use, heartbeats, completion, and classified failures without allowing collector availability to fail a build

The hosted collector and dashboard integration are in PolicyEngine/calibration-diagnostics#199.

Verification

• focused telemetry suite: 17 tests passed
• affected US release and UK build suites: passed with two expected skips
• Ruff check and format on changed Python files: passed
• repository diff check: passed