Reduce CUDA wheel size for PyPI publishing

[andimarafioti]

Summary

PyPI publishing is failing because the CUDA wheels are above the project's 100 MB per-file upload limit. The failed v0.2.0 CUDA 12.8 wheels were about 157 MiB each, almost entirely from qwentts_cpp/lib/libggml-cuda.so.0.

This PR keeps the public PyPI package on CUDA 12.8 and reduces the CUDA architecture matrix used by the publish workflow:

• CUDA 12.8 publish wheels: 75-virtual;86-real;90-real;120-real;120-virtual

The PR/dispatch wheel workflow still builds downloadable validation artifacts for CUDA 12.8, CUDA 13.0, and CPU. CUDA 13.0 artifacts use:

• CUDA 13.0 validation wheels: 75-virtual;80-real;86-real;90-real;121-real

That keeps native CUDA 13 cubins for A100 (sm_80), A10G/A10-class Ampere (sm_86), Hopper (sm_90), and DGX Spark / GB10 (sm_121), plus PTX fallback for older supported GPUs. These CUDA 13 wheels are artifacts only; they are not uploaded to PyPI.

Hardware tradeoffs for PyPI CUDA 12.8 wheels

• A10G (sm_86): native cubin, expected fast first load/TTFA.
• Hopper (sm_90): native cubin.
• CUDA 12.8 Blackwell baseline (sm_120): native cubin.
• Ada (sm_89): can use the sm_86 cubin through CUDA same-major binary compatibility.
• A100 (sm_80): PTX fallback, so first use may pay CUDA JIT cost.
• DGX Spark / GB10 (sm_121): PTX fallback in the public CUDA 12.8 wheel.

Validation

• Parsed .github/workflows/publish.yml and .github/workflows/wheels.yml with PyYAML.
• Confirmed the failing publish log reports: File too large. Limit for project 'qwentts-cpp-python' is 100 MB.
• Inspected the failed v0.2.0 artifacts:
  • x86_64 CUDA 12.8 wheel: 157.0 MiB
  • aarch64 CUDA 12.8 wheel: 156.9 MiB
  • bundled libggml-cuda.so.0: about 156.2 MiB compressed in each wheel
• Built and benchmarked CUDA 13/aarch64 comparison wheels locally on DGX Spark to verify the PTX/native first-run behavior:
  • PTX fallback: fresh-cache first TTFA 15.87s, total 16.94s; warm-cache first TTFA 466ms.
  • Native sm_121: fresh-cache first TTFA about 0.5s, total about 1.6s.
  • Broad CUDA 13 artifact policy 75-virtual;80-real;86-real;90-real;121-real: 96.49 MiB locally on aarch64, first TTFA 521ms, total 1.57s.

Notes

This should unblock PyPI without waiting for a file-size-limit increase. The package remains optimized for the CUDA 12.8 server path, while CUDA 13 wheels remain available from GitHub Actions artifacts for Spark and CUDA 13 testing.