Add portable CUDA builds for multiple NVIDIA GPU architectures

[jayzou3773]

Summary

Add a portable CUDA build and runtime compatibility layer so one LPSim release can run efficiently across common NVIDIA GPU generations, including A100 (SM80), L40 (SM89), H100/H200 (SM90), and B100/B200 (SM100).

Motivation

The repository currently defaults to CUDA architectures 80;89;90. The B200 deployment uses a separate untracked build with 90-real;90-virtual, which runs on B200 through PTX JIT rather than native SM100 code. CI only compiles SM80. This makes supported architectures, fallback behavior, and deployed artifacts inconsistent.

Proposed work

• Make the default CUDA architecture list explicit and override-friendly
• Build native cubins for SM80, SM89, SM90, and SM100
• Embed PTX fallbacks for forward compatibility
• Require a CUDA toolkit with native Blackwell compiler support for universal builds
• Add runtime reporting for GPU name, compute capability, memory, driver/runtime versions, and peer-access support
• Fail early with an actionable error for unsupported devices or an invalid GPU count
• Add a build-artifact inspection script using cuobjdump
• Document universal, H100-only, B200-only, and PTX-JIT validation commands
• Validate compilation and execution on B200 and H100 hardware
• Compare cross-GPU simulation outputs using aggregate tolerances rather than byte-for-byte equality

Acceptance criteria

• A universal build contains native SM80, SM89, SM90, and SM100 cubins
• The universal build contains documented PTX fallback code
• CUDA_DISABLE_PTX_JIT=1 runs successfully on H100 and B200
• CUDA_FORCE_PTX_JIT=1 runs successfully on H100 and B200
• Startup logs identify every selected GPU and its compute capability
• Mixed-GPU peer-access limitations are reported clearly
• Existing network, demand, and simulation output schemas remain unchanged
• Local tests and the existing GitHub Actions workflow pass

Non-goals

• Rewriting static kernels around NVRTC
• Adding architecture-specific SM100a kernels
• Guaranteeing byte-identical floating-point output across GPU generations
• Changing traffic simulation or routing algorithms