RSI: add a resumable resource-aware experiment scheduler

[w4ffl35]

Limitation\n\nRSI jobs carry resource classes and lifecycle states, but the controller does not yet schedule jobs across resource pools or execute multiple candidates concurrently. A run can checkpoint and resume, but resource arbitration is still future work.\n\n## Proposed direction\n\nAdd a durable queue/scheduler for LOCAL_GPU, CPU, REMOTE_API, and TRAINING_GPU jobs with per-resource capacity, cancellation, retry, and restart recovery.\n\n## Acceptance criteria\n\n- Resource limits prevent oversubscription across concurrent RSI runs.\n- Queued, running, completed, failed, and cancelled jobs survive restart.\n- Resume does not duplicate completed work or lose failed-job diagnostics.