Add portable CUDA architecture support

[jayzou3773]

Summary

• add toolkit-aware portable CUDA architecture profiles
• emit native cubins for A100, L40/L40S, H100/H200, and B100/B200 with PTX fallbacks
• report build/runtime GPU details and P2P availability during initialization
• move the default Docker build to CUDA 12.9.1 for native Blackwell code
• add a cuobjdump-based artifact checker, unit tests, CI coverage, and build documentation

Build behavior

CUDA toolkit	Native cubins	Embedded PTX
12.8+	sm_80, sm_89, sm_90, sm_100	compute_80, compute_100
< 12.8	sm_80, sm_89, sm_90	compute_80, compute_90

Both CMAKE_CUDA_ARCHITECTURES and LPSIM_CUDA_ARCHITECTURES remain available for custom or machine-specific builds.

Validation

Local

• python3 tests/test_cuda_artifacts.py
• demand generation and spatial partition smoke tests
• git diff --check

CUDA 12.4 compatibility build

• configured and built with the same sm_80 target used by the current GitHub Actions workflow
• verified sm_80 cubin and compute_80 PTX with cuobjdump

B200 (CUDA 12.9.1, driver API 13.2)

• portable image compiled successfully
• verified native cubins: sm_80, sm_89, sm_90, sm_100
• verified PTX: compute_80, compute_100
• full regression passed with CUDA_DISABLE_PTX_JIT=1
• full regression passed with CUDA_FORCE_PTX_JIT=1
• two-GPU probe reported bidirectional P2P availability
• average travel time: 53.86 minutes in both loading modes

H100 (Orchard job 126116, driver 580.159.03)

• ran the same portable binary produced on B200
• full regression passed with CUDA_DISABLE_PTX_JIT=1
• full regression passed with CUDA_FORCE_PTX_JIT=1
• average travel time: 53.86 minutes in both loading modes

Notes

The simulation algorithm and output schemas are unchanged. Cross-GPU validation compares invariants and numeric tolerances rather than requiring byte-identical floating-point output.

Closes #109

[Xuan-1998]

great thanks