Fix 1D GPU warning, and a few build improvements

[lgarrison]

I was confused why a user's build wasn't working, until I realized that the test code was 1D, which isn't implemented on the GPU. We were getting a low-level error message instead of our friendly error, because the check to emit the friendly error was if platform == "gpu" instead of if platform == "cuda". This PR fixes that.

While trying to diagnose this, I noticed a few other places to improve the build:

• FINUFFT_CUDA_ARCHITECTURES was using native no matter the value of CMAKE_CUDA_ARCHITECTURES, so now we propagate that value;
• FINUFFT_ARCH_FLAGS is -march=native by default, so now we document that;
• CMAKE_CUDA_ARCHITECTURES usually defaults to an old value that works but triggers PTX JIT at runtime, so now we set it to native (and document that).

I also revamped the instructions for passing CMake args to the build.