Gen 17: use --torch-backend and bump numpy to 1.26.4

[nevillelyh]

• Replace PyTorch index with --torch-backend
• Gen 17: use --torch-backend, bump numpy to 1.26.4

[nevillelyh]

❯ python3 -m monobase.diff 16 17|cut -f 2-|sort -u
certifi 2022.12.7       2025.1.31
charset-normalizer      2.1.1   3.4.1
cmake   3.25.0  4.0.0
filelock        3.13.1  3.16.1
filelock        3.13.1  3.18.0
fsspec  2024.6.1        2025.3.0
fsspec  2024.6.1        2025.3.2
hf-transfer     0.1.9   -
idna    3.4     3.10
jinja2  3.1.4   3.1.6
lit     15.0.7  18.1.8
markupsafe      2.1.5   3.0.2
networkx        3.3     3.1
networkx        3.3     3.2.1
networkx        3.3     3.4.2
numpy   1.24.1  1.24.4
numpy   1.26.3  1.26.4
numpy   2.1.2   2.2.4
nvidia-cublas-cu12      12.1.3.1        12.6.4.1
nvidia-cuda-cupti-cu12  12.1.105        12.8.90
nvidia-cuda-nvrtc-cu12  12.1.105        12.8.93
nvidia-cuda-runtime-cu12        12.1.105        12.8.90
nvidia-cudnn-cu11       9.1.0.70        9.5.0.50
nvidia-cudnn-cu12       9.1.0.70        9.5.0.50
nvidia-cufft-cu12       11.0.2.54       11.3.3.83
nvidia-curand-cu12      10.3.2.106      10.3.7.77
nvidia-cusolver-cu12    11.4.5.107      11.7.1.2
nvidia-cusparse-cu12    12.1.0.106      12.5.8.93
nvidia-nccl-cu12        2.22.3  2.26.2.post1
nvidia-nvjitlink-cu12   12.1.105        12.8.93
nvidia-nvtx-cu12        12.1.105        12.8.90
packaging       24.1    24.2
pillow  10.2.0  10.4.0
pillow  11.0.0  11.2.1
requests        2.28.1  2.32.3
setuptools      70.2.0  75.3.2
setuptools      70.2.0  78.1.0
sympy   1.13.1  1.13.3
typing-extensions       4.12.2  4.13.2
urllib3 1.26.13 2.2.3
urllib3 1.26.13 2.4.0

[nevillelyh]

warning: The --torch-backend setting is experimental and may change without warning. Pass --preview to disable this warning.
Maybe we should hold-off till this is stabilized.

[nevillelyh]

Closing as #117 is a simpler and less risky solution.