I would really appreciate it if anyone looked over the attached
patch, which adds some significant fixes to spam.el, namely:

- unregistering of articles (when they've already been registered
  with a particular spam/ham processor)

- unregistering of articles when they change from spam to ham or vice
  versa during the usage of a group summary buffer (so if you mark a
  spam message as ham or vice versa, it will be unregistered from its
  original registration) - I think I've got the algorithm right but
  any opinions are welcome

- consistent registration of articles with all the spam-use-* backends

- mass registration/unregistration, meaning that when possible a
  batch of articles will be sent to a ham/spam processor.  For tools
  like bogofilter the speed benefit should be huge, because the
  program will be invoked only once.  This will allow mass training
  and remote training eventually.

- the code is simpler, because now it's based on a few lookup tables
  instead of hand-coding all the processors.  Also, I have learned to
  abuse funcall, apply, and the ` form sometimes.

- gnus.el customizations now accept compound (list) names for
  spam/ham processors - the best way to explain it is to look at the
  gnus.el changes; the old spam/ham processors are still usable but
  marked obsolete

The fixes affect almost every part of spam.el so please look at the
patch and let me know if you see anything wrong.  I'll be using the
code in my own setup, but of course any testing you can do is much
appreciated.  I don't think the code will be bug-free because of the
scope of the fixes, that's why I'm leaving it for review before I
make a CVS commit.

Thanks
Ted