bug#64292: 30.0.50; setf strange when lexical-binding is nil

[Mattias Engdegård]

> (setq lexical-binding nil)
> (require 'nnheader)
> (macroexpand '(setf (mail-header-date header) date))
>  => (let* ((v #'(lambda (cl-x) (progn (progn (aref cl-x 3)))))
>            (v header))
>       (\(setf\ funcall\) date v v))
> 
> This looks broken, though it gets normal if lexical-binding is t.
> Because of this, an old ELisp module doesn't work.  In addition,
> the funny portion is replaced with
> \(setf\ internal--with-suppressed-warnings\)...
> if the code is byte-compiled.

This is effectively an encrypted "you should use lexical-binding:t" error 
message.
Stefan, should we just bypass the suppression here, or the entire `cl--slet` 
business, in dynbound mode?