Add RTX 50-series CUDA support

[brosequist]

Summary

• update optional PyTorch GPU requirements to CUDA 12.8 with torch 2.7.1 and torchvision 0.22.1
• update Docker build defaults to CUDA 12.8 and stop rewriting the PyTorch wheel index to CUDA 12.1
• add a CUDA architecture compatibility guard so unsupported PyTorch builds fall back to CPU with an actionable message instead of failing later with no kernel image is available for execution on the device

Context

RTX 50-series / Blackwell GPUs require PyTorch binaries with CUDA 12.8+ support. The previous optional GPU requirements installed CUDA 12.4 wheels, which can expose CUDA but fail at runtime on RTX 50-series hardware. This addresses the failure reported downstream in alicevision/Meshroom#2950.

Tests

• python3 -m py_compile segmentationRDS/segmentation.py

I did not build the Docker images locally.

[brosequist]

One heads-up for users who aren't on RTX 50-series: the cu128 wheels require CUDA runtime ≥ 12.8, which means driver ≥ 570 on Linux / ≥ 571 on Windows. RTX 30/40-series users with older drivers will need to update before using the GPU path with these requirements. Existing installs using the default CPU requirements.txt are unaffected.

[brosequist]

Quick refresh: bumped to current stable (torch==2.12.0 / torchvision==0.27.0) in 4eafdfd — the originally-pinned 2.7.1 was latest at PR-open time and is now 5 minor versions behind. CUDA 12.8 wheel index unchanged.

Checked the 24 commits on main since this PR's merge base — none of them touched torch-requirements.txt, the Dockerfiles, the build scripts, INSTALL.md, or segmentationRDS/segmentation.py, so the diff still applies cleanly with zero conflicts. The RTX 50 / CUDA 12.8 path also hasn't been addressed elsewhere in those 24 commits — the original motivation (alicevision/Meshroom#2950) still holds.

Happy to make any further changes if useful.