fused-ops: Python 3.12 support, manylinux retagging, TestPyPI toggle, wheel caching

[jilei-hao]

Summary

• Python 3.12 support: adds 3 new build matrix entries (CUDA 11.8/12.1/12.4 × Py3.12)
  via deadsnakes/ppa; drops python3.12-distutils (not available) and bootstraps
  pip with get-pip.py instead
• Manylinux retagging: renames built wheels from linux_x86_64 →
  manylinux2014_x86_64 so PyPI accepts them without a manual upload step
• TestPyPI toggle: workflow_dispatch now exposes a use_testpypi boolean input;
  the publish step routes credentials and repository URL accordingly — no more
  editing the YAML to test a release
• Wheel caching: actions/cache keyed on source file hashes skips rebuild when
  sources haven't changed, cutting CI time on re-runs
• Build tooling cleanup:
  • Switched from python setup.py bdist_wheel → python -m pip wheel (avoids
    setup.py deprecation warnings)
  • Used python -m pip consistently throughout
  • set -euo pipefail on all steps via defaults.run.shell
  • twine upload --skip-existing --verbose to survive partial re-uploads
• Dynamic package name in setup.py: name is derived from CUDA_VERSION env var
  (e.g. fireants-fused-ops-cu124); version bumped to 1.2.0; torch lower-bound
  relaxed to >=2.1.0
• Test script: scripts/test_fusedops_import_from_pypi.sh installs each
  CUDA/Torch/Python variant from TestPyPI into isolated conda envs and runs the
  import + optional pytest suite

Test plan

• Trigger publish_fusedops workflow via workflow_dispatch with
  use_testpypi: true and verify all 9 matrix jobs (Py3.10/3.11/3.12 ×
  CUDA 11.8/12.1/12.4) produce valid .whl artifacts tagged manylinux2014_x86_64
• Run scripts/test_fusedops_import_from_pypi.sh on a CUDA-capable machine
  against TestPyPI wheels to confirm import succeeds for each variant
• Trigger again with use_testpypi: false on a release tag and confirm upload
  to PyPI succeeds with --skip-existing handling duplicates gracefully