Support sm_120

[XlemonNationX]

This resolves build failures related to unrecognized CUDA architectures on systems with RTX 5090.

[pierotofy]

Thanks @XlemonNationX . This might need a few additional checks before merging as on older versions of CUDA this throws an error:

    /usr/local/cuda-11.8/bin/nvcc -forward-unknown-to-host-compiler   "--generate-code=arch=compute_70,code=[compute_70,sm_70]" "--generate-code=arch=compute_75,code=[compute_75,sm_75]" "--generate-code=arch=compute_80,code=[compute_80,sm_80]" "--generate-code=arch=compute_120,code=[compute_120,sm_120]" -MD -MT CMakeFiles/cmTC_e7f5b.dir/main.cu.o -MF CMakeFiles/cmTC_e7f5b.dir/main.cu.o.d -x cu -c /home/runner/work/OpenSplat/OpenSplat/build/CMakeFiles/CMakeScratch/TryCompile-ucUtA1/main.cu -o CMakeFiles/cmTC_e7f5b.dir/main.cu.o
    nvcc fatal   : Unsupported gpu architecture 'compute_120'
    ninja: build stopped: subcommand failed.

[XlemonNationX]

Partially revert previous changes, Fix the regular expression to only convert 2-digit architecture codes like “75” to “7.5”, and avoid incorrectly converting “120” to “1.20”, by anchoring the regex with ^ and $.

Added support for sm_120 (e.g., RTX 5090) by enabling -DOPENSPLAT_MAX_CUDA_COMPATIBILITY=ON
eg: cmake -DOPENSPLAT_MAX_CUDA_COMPATIBILITY=ON ..

[pierotofy]

Thanks @XlemonNationX !