Here is a new version of the patch. Sorry it took so long. It
addresses most of the notes I got:

Notes by Po Lu:
   - I think now the associated modifier for a key is computed
     properly, with or without XKB for plain X, and for gtk.
   - 'ralt' not generated for ISO_Level3_Shift. This will have to stay
     like this. What we know about the keys if what the system lets us.
   - Don't call x_any_window_to_frame if already called for the event.

Notes by Eli:
   - Change testing for Venable_low_level_key_events. This variable now
     is more complex.

Notes by Stefan:
   - Created and used event-* functions to access event.
   - Only one event with both pieces of information.
   - Selection of keys. Now the variable enable-low-level-key-events
     can be used to select exactly which keys we want to process.
   - About creating lisp object in handle_one_xevent, it is done in
     other events too.

The lisp interface is as follows:
   - A list of variables with keysym values is initialized (xk-*), so the
     user can select keys, not just modifiers.
   - Users can bind globally commands or functions with llk-bind to taps.
   - To use a tapped key as a new modifier, this works:
     (llk-bind 'tap 'xk-shift-r (lambda ()
                (message "H-...")
                (setq unread-command-events
                   (append (event-apply-hyper-modifier nil) nil))))
   - Added the function describe-low-level-key to see what
     keysym/modifier it triggers.


This is implemented for X, pgtkw and MS-Windows.
The pgtj implementation detects that a key is a modifier to filter it, 
but can't decide which one it its.