fix: handle amdsmi load failures when libamd_smi is missing

When the amdsmi Python package is installed without ROCm (e.g. on macOS),
import can raise KeyError while resolving libamd_smi.so. Treat AttributeError,
OSError, and KeyError like a broken install: disable AMD metrics and warn
instead of crashing the CodeCarbon CLI on import.

Author: davidberenstein1957

codecarbon/core/gpu_amd.py

@@ -28,14 +28,12 @@ def is_rocm_system():
             "Please install amdsmi to get GPU metrics."
         )
     AMDSMI_AVAILABLE = False
-except AttributeError as e:
+except (AttributeError, OSError, KeyError) as error:
     amdsmi = None
-    # In some environments, amdsmi may be present but not properly configured, leading to AttributeError when importing
     logger.warning(
-        "AMD GPU detected but amdsmi is not properly configured. "
-        "Please ensure amdsmi is correctly installed to get GPU metrics."
-        "Tips : check consistency between Python amdsmi package and ROCm versions, and ensure AMD drivers are up to date."
-        f" Error: {e}"
+        "amdsmi is installed but could not load libamd_smi (ROCm drivers missing "
+        "or incompatible). AMD GPU metrics will be unavailable. Error: %s",
+        error,
     )
     AMDSMI_AVAILABLE = False