Return NIL if the sequence is empty, instead of the NO-EXTREMUM nonsense.

 It was bad design, because it's not an error someone higher up the stack can
 sensibly handle, and handling it locally is too verbose and slow.

    (or (extremum ...) (error ...))

 expresses the common case succintly, and fits the pattern of existing
 sequence functions.

 If it is deemed necessary, we can also add &KEY DEFAULT, but that seems
 overkill and has little precedent in sequence functions.

  From http://www.cliki.net/EXTREMUM

  A simple version built on top of REDUCE for now.

  It's destructive.

 Check for empty sequences and invalid bounding index designators.

  Add bunch of missing symbols to the manual, and go over several
  docstrings reformatting them for the benefit of the documentation
  generator.

  (A crapton of symbols are still not in the manual.)

- New function: PROPER-LIST-LENGTH.

- Share code between PROPER-LIST-LENGTH, LASTCAR, and SETF LASTCAR
  via a macrolet.

- Use ENDP wrapped in SAFETY 3, to ensure proper detection of dotted
  lists.

- Use PROPER-LIST-LENGTH instead of LIST-LENGTH in ROTATE, RANDOM-ELT,
  and SHUFFLE, and test that they signal an error for improper lists.
  (Thanks to Tobias Rittweiler)

 Editing damage: extra IF surrounding the permutation guts. Also add a
 trivial test-case.

 CLHS says the third argument to DEFINE-MODIFY-MACRO must be a symbol.
 Reported by Chun Tian.

 * Thanks to Sean Ross: implement the Fisher/Yates/Knuth algorithm
   correctly.

 * As penance, specialize for lists as well: travel along the list,
   swapping towards the end -- marginally more efficient then swapping
   along the whole length.

Type declareations like ((or fixnum null) bar) or (unsigned-byte foo)
don't work on CLISP. Must use (type unsigned-byte foo) instead.