Dispatch DVM jobs to worker processes and publish results back

[Priyanshubhartistm]

Description

Follow-up to #639, #727 (job persistence), and #729 (job ingestion). The relay can now recognize NIP-90 job request events and record them as submitted jobs, but nothing actually hands them off to a worker process yet.

This issue covers dispatch: pick up a pending job, spawn/communicate with its configured worker, wait for a result (with a timeout), update job state through its lifecycle, and publish the result back as a NIP-90 event.

What needs to be done

• Pick up submitted jobs (via DvmJobRepository.findPendingJobs()) and assign them to a worke (DvmJobRepository.assignWorker(), already atomic/race-safe from feat(dvm): add dvm job persistence migration and repository #727).
• Spawn/communicate with the worker process, extending the existing spawn()-based pattern in src/cli/utils/process.ts (timeout/kill/error-classification) rather than inventing a second spawn convention.
• Frame messages to/from the worker as newline-delimited JSON over stdin/stdout no existing convention covers this (process.ts buffers all output into one string, which doesn't work for a long-lived worker handling multiple in-flight jobs), so this needs new code.
• On timeout: mark the job timed_out via DvmJobRepository.updateStatus() and kill the worker process (soft Node-level guard — timeout + child.kill(), not full OS-level sandboxing; that's a stretch goal, not required for this).
• On success: mark the job completed, publish a kind 6000-6999 job result event.
• On worker error/crash: mark the job failed with an error message.
• Tests for the dispatch flow: successful completion, timeout, worker crash, and the state transitions in each case.