[SWDEV-477447] Set _HAS_PYNVML to false if amdsmi not installed (#1535)

Fix from pytorch#132990 cherry picked
into 6.2 to be resolved for 6.2.1

"""
This is a bugfix that was recently encountered in ROCm/Deepspeed.
Currently if a library installs pynvml and runs on ROCm pytorch will
break as _HAS_PYNVML is set to true and it will attempt to use amdsmi
library for the device_count call which will not be installed.

This fix will set _HAS_PYNVML to false on ROCm if amdsmi is not
installed.
"""

Will need to be cherry picked to release/2.3 and
rocm6.3_internal_testing also

Author: jataylo

torch/cuda/__init__.py

@@ -54,13 +54,10 @@
 _PYNVML_ERR = None
 try:
     try:
-        import pynvml  # type: ignore[import]
-
-        _HAS_PYNVML = True
-    except ModuleNotFoundError:
-        pass
-    try:
-        import amdsmi  # type: ignore[import]
+        if not torch.version.hip:
+            import pynvml  # type: ignore[import]
+        else:
+            import amdsmi  # type: ignore[import]
 
         _HAS_PYNVML = True
     except ModuleNotFoundError: