Faster attach

[plicease]

I think that calling an attached function could be a lot faster if we did something like what FFI::Platypus does with its attach (using the any_ptr trick).

We should: create an optional XS module that handles performance critical parts of the API where it makes sense, like calling attached functions (the dynamic call and callback are two other candidates). I think it would be counter productive and pointless to reimplement the entire API in XS.

TheXS module should only be released if the benchmarks prove it useful.