CUDA_Image_Selection_for_Build_&_Runtime

[lost-particles]

Introduced build/runtime CUDA flavor args with optional build stage for wheels
Fixes issue #1288

🚀 What

• Introduced BUILD_CUDA_FLAVOR and RUNTIME_CUDA_FLAVOR arguments (default base)
• Added lightweight build stage to pre-build wheels when using devel
• Kept final runtime image defaulting to base while allowing flexible overrides

💻 How

• Added two ARGs to dockerfile.jinja and used them in the FROM statements
• Created a minimal build stage to compile Python/CUDA wheels when needed
• Updated requirements installation to prefer prebuilt wheels from the build stage

🔬 Testing

• Built the image with default args to confirm base/base flow remains unchanged
• Built with --build-arg BUILD_CUDA_FLAVOR=devel to confirm CUDA ops compile correctly and runtime stays base
• Built with --build-arg BUILD_CUDA_FLAVOR=devel --build-arg RUNTIME_CUDA_FLAVOR=devel to confirm optional devel runtime works as expected

[bolasim]

Thanks for the PR! Have you considered using base_image instead of needed to add more variants of the default base images?