Document RTX 50-series CUDA 12.8 setup

[Lucas1479]

Summary

• document the RTX 50-series / Blackwell incompatibility with the current torch==2.5.1+cu124 profile
• provide the community-validated PyTorch 2.7.0 CUDA 12.8 installation command in Chinese and English
• keep CUDA 12.4 as the official baseline and state that the CUDA 12.8 combination has not yet passed the full project regression gates

Validation

• git diff --check
• documentation-only change

[Lucas1479]

Maintainer self-review complete: no blocking findings. PyTorch 2.7 officially introduced Blackwell support with CUDA 12.8 wheels (https://pytorch.org/blog/pytorch-2-7/), and the official previous-versions page lists the exact torch 2.7.0 / torchvision 0.22.0 / torchaudio 2.7.0 cu128 command for Linux and Windows (https://docs.pytorch.org/get-started/previous-versions/#v270). The live cu128 package index still exposes compatible Python 3.12 Windows wheels for all three versions. Self-review added one safety clarification: run the override only inside a separate activated experimental venv such as .venv_cu128, never by overwriting the formal .venv_cu124 environment. The README continues to label cu124 as the official project baseline and cu128 as not yet fully regression-qualified. Diff check, provenance (0 errors), deterministic source-release (0 errors/0 warnings), and both required build/cpu-model-less checks passed on final head 6b51bab. Admin merge bypass is used only for the unavailable external approval requirement.