Fix torch_tensorrt import crash on non-GPU nodes

torch_tensorrt now raises RuntimeError (not ImportError/OSError) when
no CUDA device is available. Add RuntimeError to the except clause so
convnet loads gracefully, allowing mazepa_layer_processing and other
modules to register their builders on non-GPU headnodes.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: supersergiy

zetta_utils/convnet/utils.py

@@ -24,7 +24,7 @@
     import torch_tensorrt  # pylint: disable=import-error
 
     TENSORRT_AVAILABLE = True  # pragma: no cover
-except (ImportError, OSError) as e:
+except (ImportError, OSError, RuntimeError) as e:
     logger.info(f"torch_tensorrt is not available: {e}")