Remove layer_norm from TPU benchmark (OOB slice bug, gh#1937)

Author: norx1991

.github/workflows/benchmark_tpu_nightly.yml

@@ -16,7 +16,8 @@ on:
         #   rms_norm: InductorLoweringError in torch.mean reduction codegen for fori_loop/emit_pipeline
         #   geglu/swiglu: autotuning takes >15min per kernel (large shape 8x2048x4096), many configs fail to compile
         #   low_mem_dropout: ~37% element accuracy mismatch on all configs except block_sizes=[128]
-        default: "exp,add,softmax_two_pass,welford,layer_norm"
+        #   layer_norm: OOB slice when reduction_loops doesn't evenly divide the reduction dim (gh#1937)
+        default: "exp,add,softmax_two_pass,welford"
 
 permissions:
   contents: read
@@ -27,4 +28,4 @@ jobs:
     permissions:
       contents: read
     with:
-      kernels: ${{ github.event.inputs.kernels || 'exp,add,softmax_two_pass,welford,layer_norm' }}
+      kernels: ${{ github.event.inputs.kernels || 'exp,add,softmax_two_pass,welford' }}

benchmarks/run_tpu.py

@@ -107,7 +107,6 @@ def _softmax_shapes() -> list[tuple[str, tuple[Any, ...]]]:
         _softmax_shapes,
     ),
     "welford": ("welford", "welford", None, None),
-    "layer_norm": ("layer_norm", "layer_norm", None, None),
 }