Update cuda support gpu architecture

[LambdaGo-lab]

What changed

• Updated CMAKE_CUDA_ARCHITECTURES logic to correctly handle CUDA 13.0+.
• Removed GPU architectures that are no longer supported by CUDA 13.0 (Maxwell, Pascal, Volta).
• Kept Turing (sm_75) and newer architectures enabled (Ampere, Ada, Hopper).

Why this is needed

CUDA 13.0 dropped support for offline compilation targeting:

• sm_50, 52, 53 (Maxwell)
• sm_60, 61, 62 (Pascal)
• sm_70, 72 (Volta)

Keeping these architectures in the build causes nvcc errors or requires deprecated flags.
This change aligns the CMake configuration with NVIDIA’s supported architecture set for CUDA 13.0+.

Supported architectures after this change (CUDA 13.0+)

• sm_75 (Turing)
• sm_80, sm_86, sm_87 (Ampere)
• sm_89 (Ada Lovelace)
• sm_90 (Hopper)

Backward compatibility

• CUDA 12.x and 11.x behavior is unchanged.
• Older architectures remain enabled when using older CUDA toolkits.

Notes

• -Wno-deprecated-gpu-targets is preserved to avoid warnings on older toolchains.