fix: stabilize CUDA build and validation

[LessUp]

Summary

• stabilize the local CUDA/CMake toolchain defaults and architecture handling for nvcc 12
• fix Tensor Core fallback/test source issues and add a shared CUDA-aware gtest environment
• correct theoretical peak GFLOPS math and make performance regression thresholds realistic for the current teaching kernels

Validation

• cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
• cmake --build build -j$(nproc)
• ctest --test-dir build --output-on-failure
• npm install -g @fission-ai/openspec@latest && openspec validate --all
• npm --prefix docs ci && npm --prefix docs test && npm --prefix docs run build