Thank you for your excellent work on `sort` and related functionality!

Unfortunately, the new `sort` implementation occasionally crashes with a
segfault. The following code reproduces that in current master:

(dotimes (i 500)
  (sort (make-list 128 42)
        :key (lambda (n) (make-list i n))))

It happens for inputs of length >= `MERGESTATE_TEMP_SIZE / 2` (= 128
currently) along with a non-NIL `:key` function. In such cases, a
`Lisp_Object` array is explicitly heap-allocated to store the keys, which is
never marked against GC. This would not be a problem if not for the fact that
the `:key` function call may trigger GC.

I'm attaching a patch with a proposed solution, consisting of three changes:

1. Allocate with `xzalloc` to have the new array initialized to `Qnil`. This
   ensures its objects can be marked properly.

2. Mark `ms->allocated_keys` in `merge_markmem`. We don't need to check that
   `ms->allocated_keys != NULL`, as `merge_register_cleanup` is only called
   under this exact condition.

3. Move the computation of keys (which may trigger a GC) after `merge_init`,
   which registers `merge_markmem`.

I hope this is useful.

Cheers,