If 'all-completions' is called under certain conditions,
case-fold-search is specbound and unbound for each matching candidate.
Because this variable is DEFVAR_PER_BUFFER, specbinding it is slow
(scales with number of buffers).  This patch eliminates specbinding from
the three core completion functions.

Benchmark:

  (dotimes (i 300)
    (get-buffer-create (format " *test-buffer-%s*" i)))

  (let ((completion-regexp-list '("\\`.*?")))
    (benchmark-run-compiled 50
      (all-completions "" obarray #'boundp)))

9.9 seconds without patch,
0.83 seconds with patch applied.

Note that for the performance issue to be observed, we must have a lot
of live buffers, completion-regexp-list must be non-nil and a predicate
must be passed to all-completions.  The last two conditions are
satisfied if we press M-x TAB.