[Power] support dedicated infrastructure with measured GPU power - #424

Power telemetry currently rejects a dedicated etcd/NATS node because its placement moves the benchmark away from the collector's clock. Keep the actual Slurm batch host as head and benchmark host, and reserve another node for infrastructure. This supports homogeneous and heterogeneous allocations while excluding the infrastructure node from serving GPUs.

Configuration validation rejects dedicated frontend/client nodes and srun_options node-placement overrides when power is enabled. Dry-run output and the configuration reference describe the placement. This uses the existing collector and applies to fixed-sequence power workloads; it does not add AgentX collection support.

Testing

• python -m pytest tests/test_configs.py tests/test_telemetry.py tests/test_dry_run.py: 369 passed, covering shared/dedicated infrastructure, both allocation types, batch-host placement, worker exclusion, invalid overrides, and dry-run output.
• ruff check src, source and changed-test formatting checks, and git diff --check: passed.
• Earlier make check: 1,950 passed and four failures. Two new dry-run assertions were corrected for Rich table line wrapping and pass in the final focused run. The other two failures reproduce on untouched base ed72cbef: the mock submission assertion in test_apply_mock.py and the CPU-affinity test in test_fingerprint.py (os.sched_getaffinity is absent on macOS). The repository's non-blocking type check emitted diagnostics. The full suite was not repeated after the assertion fix.
• CPU-only validation; no Slurm/GPU run or CI dispatch.