Capture a running sandbox as a new forkable snapshot [rate class: sandbox_api_rps]

Capture a running sandbox's CURRENT state - memory, processes, disk - as a new forkable snapshot, without stopping the sandbox. This is how a computed prefix gets shared instead of recomputed: run the common steps once in one sandbox, snapshot_sandbox, then fork_sandbox that snapshot so every branch starts from the finished work. Use it at branch points (explore K continuations of one trajectory), before risky steps (retry from the checkpoint instead of from scratch), and to compare approaches from a truly identical starting state. Snapshots are content-addressed against their parent, so capturing one is cheap and far faster than re-running the steps it contains.