feat(research): support CUDA convergence runs

Allow the sealed ResNet and YOLO convergence protocol to run unchanged on the remote NVIDIA accelerator while retaining CPU and MPS support.

Constraint: Preserve fixed seeds, exact workloads, and one-device-per-run semantics

Rejected: Reuse mixed MPS and CUDA results | invalid cross-device production matrix

Confidence: high

Scope-risk: narrow

Not-tested: Full production matrix on RTX 3080 Ti

test/post_training_resnet_convergence.py

@@ -790,8 +790,10 @@ class ResNetConvergenceAdapter:
         self.data_root = Path(data_root)
         self.device = torch.device(device)
         self.allow_download = bool(allow_download)
-        if str(self.device) not in {"cpu", "mps"}:
-            raise ProtocolError("ResNet device must be cpu or mps")
+        if str(self.device) not in {"cpu", "mps", "cuda"}:
+            raise ProtocolError(
+                "ResNet device must be cpu, mps, or cuda"
+            )
 
     def _result_path(self) -> Path:
         return self.run_root / "workloads" / self.workload_id / "result.json"