diff.c: fix a double-free regression in a18d66cefb

My a18d66cefb9 (diff.c: free "buf" in diff_words_flush(), 2022-03-04) has what it retrospect is a rather obvious bug (I don't know what I was thinking, if it all): We use the "emitted_symbols" allocation in append_emitted_diff_symbol() N times, but starting with a18d66cefb9 we'd free it after its first use! The correct way to free this data would have been to add the free() to the existing free_diff_words_data() function, so let's do that. The "ecbdata->diff_words->opt->emitted_symbols" might be NULL, so let's add a trivial free_emitted_diff_symbols() helper next to the function that appends to it. This fixes the "no effect on show from" leak tested for in the preceding commit. Perhaps confusingly this change will skip that test under SANITIZE=leak, but otherwise opt-in the "t4015-diff-whitespace.sh" test. The reason is that a18d66cefb9 "fixed" the leak in the preceding "no effect on diff" test, but for the first call to diff_words_flush() the "wol->buf" would be NULL, so we wouldn't double-free (and SANITIZE=address would see nothing amiss). With this change we'll still pass that test, showing that we've also fixed leaks on this codepath. We then have to skip the new "no effect on show" test because it happens to trip over an unrelated memory leak (in revision.c). The same goes for "move detection with submodules". Both of them pass with SANITIZE=address though, which would error on the "no effect on show" test before this change. Reported-by: Michael J Gruber <git@grubix.eu> Signed-off-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com> Signed-off-by: Junio C Hamano <gitster@pobox.com>