Clean up and optimize self-contained LayerNorm backward for GB300

[Laurawly]

Primary public benchmark mode is now CUDA graph warm replay:

  env PYTHONNOUSERSITE=1 CUTE_DSL_ARCH=sm_103 PYTORCH_ALLOC_CONF=expandable_segments:True \
    conda run -n cute python -u oink/benchmarks/benchmark/benchmark_layernorm_bwd_sm100.py \
      --dtype bf16 --weight-dtype same --dsv3 --iters 80 --warmup-ms 10 --cuda-graph \
      --json /tmp/oink_layernorm_bwd_sm103_dsv3_cuda_graph_seq.json

  env PYTHONNOUSERSITE=1 CUTE_DSL_ARCH=sm_103 PYTORCH_ALLOC_CONF=expandable_segments:True \
    conda run -n cute python -u oink/benchmarks/benchmark/benchmark_layernorm_bwd_sm100.py \
      --dtype bf16 --weight-dtype same --dsv4 --iters 80 --warmup-ms 10 --cuda-graph \
      --json /tmp/oink_layernorm_bwd_sm103_dsv4_cuda_graph_seq.json

  DSv3 LayerNorm backward, bf16/same/no-bias, CUDA graph replay

       M      N   Oink ms   Oink TB/s   ATen ref ms   Oink/ref
  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
    4096   6144    0.0548      2.7574        0.0777    1.4190x
    4096   8192    0.0611      3.2951        0.0970    1.5873x
   16384   6144    0.1840      3.2833        0.2794    1.5183x
   16384   8192    0.2093      3.8480        0.3387    1.6183x
   65536   6144    0.6896      3.5043        1.0652    1.5447x
   65536   8192    0.7372      4.3705        1.3138    1.7823x

  DSv4 hidden LayerNorm backward, bf16/same/no-bias, CUDA graph replay

       M      N   Oink ms   Oink TB/s   ATen ref ms   Oink/ref
  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
    4096   7168    0.0591      2.9800        0.0858    1.4503x
   16384   7168    0.1990      3.5425        0.3077    1.5467x
   65536   7168    0.7467      3.7753        1.1711    1.5684x