Replace individual ROCm libs with rocm-libs meta-package; graceful provider fallback

- Dockerfile: rocm-libs installs all ROCm compute libraries at once
  (hipblas, hipfft, hipsparse, rocblas, miopen, etc.) avoiding whack-a-mole
  with individual missing .so files
- engine.py: query ort.get_available_providers() at runtime and only request
  providers that actually loaded — falls back to CPU instead of crashing

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

engine.py

@@ -28,7 +28,8 @@ _cond_cache: Dict[str, Dict[str, np.ndarray]] = {}
 def detect_device() -> str:
     try:
         import onnxruntime as ort
-        if "ROCMExecutionProvider" in ort.get_available_providers():
+        available = ort.get_available_providers()
+        if "ROCMExecutionProvider" in available or "MIGraphXExecutionProvider" in available:
             return "rocm"
     except Exception:
         pass
@@ -36,13 +37,19 @@ def detect_device() -> str:
 
 
 def _get_providers(device: str) -> list:
-    if device in ("rocm", "cuda"):
-        return [
-            "MIGraphXExecutionProvider",
-            ("ROCMExecutionProvider", {"device_id": 0}),
-            "CPUExecutionProvider",
-        ]
-    return ["CPUExecutionProvider"]
+    if device not in ("rocm", "cuda"):
+        return ["CPUExecutionProvider"]
+
+    import onnxruntime as ort
+    available = set(ort.get_available_providers())
+    providers = []
+    if "MIGraphXExecutionProvider" in available:
+        providers.append("MIGraphXExecutionProvider")
+    if "ROCMExecutionProvider" in available:
+        providers.append(("ROCMExecutionProvider", {"device_id": 0}))
+    providers.append("CPUExecutionProvider")
+    logger.info(f"Available ORT providers: {available}")
+    return providers
 
 
 def load_model() -> bool: