fix(cloud-eval): lift grader execution errors into RowMetric.error

[placerda]

When a Foundry azure_ai_evaluator grader fails to execute (most commonly because the evaluator service principal lacks Cognitive Services OpenAI User on the target model deployment), the per-metric score comes back null and the real cause is buried in result.sample.error.message. Operators see only actual=missing in the threshold table and the CI gate fires as threshold_failed for the wrong reason.

Root cause (real shape from PO's run)

{
  "name": "coherence",
  "score": null,
  "passed": null,
  "status": "error",
  "sample": {
    "error": {
      "code": "FAILED_EXECUTION",
      "message": "OpenAI API hits AuthenticationError: 401 PermissionDenied: The principal lacks the required data action chat/completions/action"
    }
  }
}
```nThe parser only probed `error` as a top-level string. It never reached `sample.error.message`.

## Fix

- `cloud_results._metric_from_result` now calls a new `_extract_grader_error` that probes (in order): top-level `error`, `sample.error`, and `status == 'error'` as last resort. Accepts both string and `{code, message}` dict shapes; prefixes the error code when present.
- `orchestrator` 0-usable-scores warning now quotes the first per-metric error so CI logs carry the actionable cause without operators having to download the raw artifact.
- 3 new unit tests in `test_cloud_results.py` covering: `sample.error.message` dict lift, top-level `error` dict, and the case where a real score must not be shadowed by a stale `sample.error: null`.

## Verification

- `python -m pytest tests/unit/test_cloud_results.py -x -q` → 14 passed.
- `python -m pytest tests/ -x -q` → 792 passed, 3 skipped.
- Real artifact from `placerda/agentops-prompt-quickstart` run #26618645299 shown above as the source of the fixture.