Fix CUDA index selection for pinned torch versions

[godnight10061]

Problem

When users pin PyTorch 2.7.x (or torchvision 0.22.x / torchaudio 2.7.x), torchruntime may still select the cu124 index URL based on detected hardware/Python version.

PyTorch 2.7 wheels are published under cu128, so pip fails with No matching distribution found when it is pointed at the cu124 index.

Conversely, if hardware defaults to cu128 but the caller pins torch <=2.6, those wheels live under cu124 and pip fails for the same reason.

This is the root cause of #16.

Solution

• Infer the required CUDA wheel index (cu124 vs cu128) from pinned torch / torchvision / torchaudio requirements passed to torchruntime.install().
• If the inferred index conflicts with the detected CUDA platform, override only the cu124/cu128 portion so the correct https://download.pytorch.org/whl/<platform> index URL is used.
• Preserve the nightly/ prefix when applicable.
• Do not demote Blackwell (arch 12) from cu128 -> cu124 since older torch versions won’t support it anyway.

If the pinned requirements imply conflicting indices, torchruntime leaves the platform unchanged and lets pip resolve/fail normally.

Tests

• Added regression tests that fail on main and pass with this change.
• python -m pytest (141 tests) passes.

Fixes #16.