Restore mtlcall

[maleadt]

We used to have the ability to call pre-compiled kernels (e.g. by passing Metal source code to the appropriate API functions), see https://github.com/JuliaGPU/Metal.jl/tree/9afb62460f8005db00dd3ea71a278758853b24e9/examples/driver. That got lost when we started relying on argument metadata generated by GPUCompiler to configure argument encoders. Ideally we'd still retain the ability to use non-Julia kernels, but it's not clear how to do that (do we want to be able to reconstruct the argument metadata from a tuple type? or do we only allow 'simple' arguments for such kernels, i.e., that do not require argument encoders?).