feat(telemetry): MVP support for opt-in telemetry of runner crashes

[AndreiCravtov]

Motivation

Runner crashes currently leave useful diagnostics only on the local machine. This adds an opt-in path for collecting raw runner stderr logs so crash reports can be inspected centrally without making telemetry part of the critical execution path.

Changes

• Added --telemetry, disabled by default.
• Added EXO_TELEMETRY_API_URL, defaulting to https://telemetry.exolabs.net/.
• Added TelemetryService and TelemetrySink for bounded, non-blocking telemetry submission.
• Wired telemetry from Node through Worker into RunnerSupervisor.
• Changed runner stderr logging to write per-runner timestamped files under:
  runner_log/<instance_id>/<runner_id>/<utc_timestamp>.stderr.log
• Upload runner stderr logs by:
  • reading the completed stderr file
  • skipping zero-byte files
  • hashing the raw file bytes with SHA-256
  • requesting a pre-signed upload URL
  • uploading the raw bytes to that URL
• Swallow telemetry upload failures after logging a warning.

Why It Works

Telemetry is isolated behind a bounded channel and best-effort background service, so failures, slow uploads, or backpressure do not block runner supervision or node operation.

Runner stderr files are written per bound instance/runner with UTC timestamped names, avoiding the previous shared append-only log path. Hashing the final file contents before upload gives the backend a stable object identity and lets it verify that uploaded contents match the requested submission.

Test Plan

Manual Testing

• Ran exo with --telemetry against the telemetry API.
• Forced a runner failure and confirmed a non-empty runner stderr log was submitted through the pre-signed upload flow.

Automated Testing

• Added unit coverage for runner stderr telemetry upload:
  • SHA-256 presign payload generation
  • raw byte upload to the returned URL
  • upload failure handling without propagating exceptions
• Updated runner supervisor tests to use isolated per-test log paths and a dummy telemetry service.