Since libffi introduced static trampolines in commit libffi/libffi@9ba5592 , it's necessary to use ffi_prep_cif() and ffi_prep_closure_loc() instead of our own optimized trampoline. Otherwise all functions attached to a FFI::Library module result in a segfault.

Using ffi's (classic or static) trampolines is somewhat slower than our optimized ones, but they are still faster than the alternative ruby code like this:

module FFI
  class Function < Pointer
    def attach(mod, name)
      this = self
      body = -> *args, &block do
        this.call(*args, &block)
      end
      mod.define_method(name, body)
      mod.define_singleton_method(name, body)
      self
    end
  end
end

Fixes #989