Remove ban on using string_cast with CUDA host code
string_cast.hpp merely detects whether the current compiler is NVCC (originally based on `if defined(__CUDACC__)` in glm/simd/platform.h) and throws an error if it is. This means string_cast.hpp cannot be included in any header which might ever be used in a CUDA project.
Of course, glm::to_string can't be used in device (GPU) code. However, the current approach to stop this is both incorrect and unnecessary. __CUDACC__ will be defined in both host and device code compilation, and glm::to_string can obviously be used in host code. The correct define is __CUDA_ARCH__ (will be defined only if compiling device code). However, there's no problem if glm::to_string is defined (the header is included) while compiling device code, as long as it's not actually used in the device code. So, throwing an error if __CUDA_ARCH__ is defined would still prevent string_cast.hpp from being included in CUDA projects.
There's actually no need for any manual check to see if glm::to_string is being used in device code, because the compiler will already check for that. It returns a std::string, which itself can't be used in device code, so it's unlikely a developer would try. And if they did, there would be errors that both glm::to_string and all the needed std::string constructors, stream operators, etc. are host-only functions.
lpisha