Support Blackwell GPUs in setup

[Arison591]

Summary

Addresses #2.

• detect the highest NVIDIA compute capability, with ANIGEN_CUDA_CAPABILITY as a fallback for environments where GPU discovery is unavailable
• select PyTorch 2.7.0 with the CUDA 12.8 wheel index on Blackwell GPUs, while preserving the existing PyTorch 2.4/cu121 selection for earlier CUDA 12 GPUs
• export TORCH_CUDA_ARCH_LIST and FLASH_ATTN_CUDA_ARCHS for local extension builds and reject PyTorch versions older than 2.7 on Blackwell
• document the Blackwell path and add a self-contained installer regression test

PyTorch 2.7 is the first release with Blackwell support and CUDA 12.8 wheels: https://pytorch.org/blog/pytorch-2-7/

Why

The current CUDA-major-only selection installs PyTorch 2.4.0 from the cu121 index for every CUDA 12 GPU. That wheel does not contain support for the RTX 50-series sm_120 target, so the default setup path is not usable on Blackwell.

Validation

• tests/test_setup_blackwell.sh — passed (Blackwell selection, legacy selection, and old-Torch rejection)
• bash -n setup.sh tests/test_setup_blackwell.sh — passed
• shellcheck -e SC2166 setup.sh tests/test_setup_blackwell.sh — passed (SC2166 is from a pre-existing conditional)
• NVIDIA GeForce RTX 5090, compute capability 12.0, driver 580.76.05:
  • PyTorch 2.7.0+cu128 SDPA CUDA forward — passed with finite output
  • spconv-cu121==2.3.8 minimal SubMConv3d CUDA forward — passed, output shape (4, 5)
  • flash-attn==2.8.3 flash_attn_func CUDA forward — passed with finite output, input shape (2, 128, 8, 64)

I did not download the multi-GB AniGen checkpoints or run a full mesh-generation workflow.