Prototype CompilerPlugins in GPUCompiler 😱

[vchuravy]

Mashup of #582 and #599 with a heavy dose of JuliaLang/julia#52964

Goal is to enable:

function myfma(a,b,c)
          a*b+c
end
function kernel(ptr, a, b, c)
           unsafe_store!(ptr, CCMixin._call_within(FMACompiler(), myfma, a,b,c))
           return
end

But I am currently struggling with inlining and I am taking a detour to first get the LLVM emission right...