Excessive compile times and CUDA binary size due to excessive use of constexpr

[CommanderLake]

Excessive use of constexpr is multiplying the compiled code paths and producing a ggml-cuda.dll binary that can now exceed 400MB for little to no performance benefit at the cost of excessive compile times, download times, required storage space and memory just to load the dll!

Reduce unnecessary usage of constexpr, or better yet add an option to reduce constexpr usage to speed up compile times and reduce the compiled binary size.