The MXFP8 nested-reduction test directly called the private
cvt_e8m0_rceil prim, whose lowering only supports NVIDIA SM100+. This
made the test fail on XPU and expanded the proposed fix into an
unnecessary cross-device lowering.

Use Inline ASM HOP on CUDA and ordinary tensor bit operations on XPU.
This keeps the direct prim unchanged while exercising the same fused
nested-reduction kernel on both backends.

Test Plan:

```
bash -lc 'source "/home/xu/intel/oneapi/setvars.sh" --force >/dev/null 2>&1 && conda run -n repro_194562_source python "/home/xu/conda_root/xu_pytorch/fix_torch_xpu_ut/xuhancn_src/pytorch/test/inductor/test_nested_reduction.py" -k test_rmsnorm_mxfp8_scale_swizzle_kernel_form -v'
PATH="/home/xu/miniforge3/bin:/home/xu/.local/bin:/home/xu/.local/bin:/home/xu/bin:/home/xu/.opencode/bin:/home/xu/bin:/home/xu/miniforge3/bin:/home/xu/miniforge3/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin" conda run -n pr193554_lint spin quicklint
PATH="/home/xu/miniforge3/bin:/home/xu/.local/bin:/home/xu/.local/bin:/home/xu/bin:/home/xu/.opencode/bin:/home/xu/bin:/home/xu/miniforge3/bin:/home/xu/miniforge3/condabin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/games:/usr/local/games:/snap/bin" conda run -n pr193554_lint lintrunner -a test/inductor/test_nested_reduction.py
```

Authored with assistance from an AI assistant.