Prevent Numba Performance Warnings in cuDF (#9320)

Numba 0.54 introduces performance warnings whenever a kernel is launched
with low occupancy. Certain operations in cuDF can indirectly cause this
warning to be emitted, for example `DataFrame.iloc` (which one would not
expect to be particularly efficient anyway). This message could be
disconcerting for users and appear in a lot of workflows; therefore,
this commit configures Numba at cuDF import time to suppress the low
occupancy warning.

Fixes #9247.

python/cudf/cudf/__init__.py

@@ -4,7 +4,7 @@
 validate_setup()
 
 import cupy
-from numba import cuda
+from numba import config as numba_config, cuda
 
 import rmm
 
@@ -104,5 +104,13 @@
 cuda.set_memory_manager(rmm.RMMNumbaManager)
 cupy.cuda.set_allocator(rmm.rmm_cupy_allocator)
 
+try:
+    # Numba 0.54: Disable low occupancy warnings
+    numba_config.CUDA_LOW_OCCUPANCY_WARNINGS = 0
+except AttributeError:
+    # Numba < 0.54: No occupancy warnings
+    pass
+del numba_config
+
 __version__ = get_versions()["version"]
 del get_versions