Implement durable execution with automatic recovery from checkpoints

[lanycrost]

Summary

Add checkpoint-based execution so workflows can be resumed from the last checkpoint if a pod dies mid-execution.

Problem

Currently there is no automatic replay from checkpoint. If a pod dies mid-workflow, the entire step must be retried. Users must rely on external state stores and step-level retries.

Proposed change

• Checkpoint API for steps to persist intermediate state
• Controller-driven recovery that resumes from the last checkpoint
• Integration with existing storage offloading infrastructure

References

• Roadmap: https://bubustack.io/docs/community/roadmap (Known gaps, Near-term vision)
• Durable semantics: https://bubustack.io/docs/overview/durable-semantics