fix: solo/self-hosted deployments inherit expensive sandbox defaults

[mlekhi]

Self-hosting a solo instance burned ~12h of my Active CPU over ~5 days of no use. Usage was zero on the days before it was deployed, so it's not a leak. Fluid CPU kept growing on days I never opened the app, which points to something running on its own. Possible issues could be:

• Likely root cause: stale activeStreamId only checks the ID is non-null, not that the run is alive. If it's orphaned, the lifecycle workflow (sandbox-lifecycle.ts:95) never hibernates and re-loops every 5s.
• Eager provisioning on session create (sessions/route.ts:408, Start sandbox provisioning on session create #880) - abandoned sessions still spin up a sandbox. That + 4 vCPUs per sandbox is significant.
• 30-min idle window (config.ts:31) - sandboxes stay alive 30 min after last activity.

Suggestions (I'll get to these when I have codebase access lol)

• Check run status in hasActiveStreamForSession instead of just non-null, and clear activeStreamId server-side on workflow completion
• Document OPEN_AGENTS_RESOURCE_PROFILE=hobby env var in README / .env.example
• Make provisioning lazy (provision on first message)
• Lower / make configurable the idle window