fix: sb polling fixes

[luke-lombardi]

Summary by cubic

Stabilizes sandbox polling and restore. Makes process manager readiness reliable, fixes restore timing, improves exec streaming/logs, and auto-creates local S3 buckets in dev.

• Bug Fixes

  • Process manager readiness: TCP probe + Ready RPC with backoff; try container IP then published address; classify/transient errors; evict and retry client; record failure stats; cache container address map on the instance.
  • Exec streaming: stream sandbox exec output; ack log chunks after writing; handle early stream close; surface early-exit errors; keep logs buffered on event write failure.
  • Process listing: wait for the manager; add /proc scan fallback for restored sandboxes to include missing PIDs.
  • Restore flow (gVisor): wait for runtime state and only signal Started after PID; nudge restored process manager with SIGWINCH; clean up restore on failure/cancel; worker polls restored runtime until it exits.
  • runc: map “not found” to a typed error.
  • Docker-in-Docker: treat user stop or teardown as cancel during startup; run cgroup setup and IPv4 enable via streamed exec.
  • Storage/Gateway: add EnsureBucket; ensure local S3 buckets exist for dev endpoints on object head/create.
  • Sandbox status/API: ping runtime for running containers; evict client and retry on transient connect; return stable pending/stopping/running states.
  • SDK: retry exec only on explicit readiness errors, including “process manager failed to become ready”.

• Dependencies

  • Bump github.com/beam-cloud/goproc to v0.1.11.

^{Written for commit 5f1cf5a. Summary will update on new commits.}