feat: add spinloop work abort #208

Stop a running work item and return it to the backlog: the item's agent is stopped cleanly and its record is removed from the state, so the orchestrator's next pass admits it again.

• Takes --items <path> (default ./work.yaml), the same file the orchestrator reads
• Only a running item can be aborted; aborting a backlog, done, or failed item is a refusal naming the item and its state
• With no orchestrator running there is nothing running: the command says so. Stale running records from an unclean death are for the next start to record failed, not for abort to clear
• The orchestrator owns the state file and holds the lock, so how the abort reaches it — a marker beside the file, a control channel — is a design decision for the change. The requirement is the outcome: the agent stops, the item is back in the backlog

Part of the work command family with #206, #207, and #209. Builds on #205.