Better way to handle SIMD

[jgraef]

simd.rs handles detecting SIMD capabilities and provides some methods to query how certain vector and matrix types should be vectorized. This still leads to a lot of code duplication, because often the codegen between different cases is still very similar. E.g. adding 2 vectors just needs to generate iadd for every pair IR values. This is somewhat avoided with the try_map_as_scalars and try_zip_map_as_scalars methods, but they just pretend the vector IR types are just scalars. A better abstraction would be nice.