alert_words: Fix case-sensitivity of alert words.

Previously, alert words were case-insensitive in practice, by which I
mean the Markdown logic had always been case-insensitive; but the data
model was not, so you could create "duplicate" alert words with the
same words in different cases.  We fix this inconsistency by making
the database model case-insensitive.

I'd prefer to be using the Postgres `citext` extension to have
postgres take care of case-insensitive logic for us, but that requires
installing a postgres extension as root on the postgres server, which
is a pain and perhaps not worth the effort to arrange given that we
can achieve our goals with transaction when adding alert words.

We take advantage of the migrate_alert_words migration we're already
doing for all users to effect this transition.

Fixes #12563.

zerver/tests/test_alert_words.py

@@ -56,6 +56,16 @@ class AlertWordTests(ZulipTestCase):
         words = user_alert_words(user)
         self.assertEqual(set(words), set(self.interesting_alert_word_list))
 
+        # Test the case-insensitivity of adding words
+        add_user_alert_words(user, set(["ALert", "ALERT"]))
+        words = user_alert_words(user)
+        self.assertEqual(set(words), set(self.interesting_alert_word_list))
+
+        # Test the case-insensitivity of removing words
+        remove_user_alert_words(user, set(["ALert"]))
+        words = user_alert_words(user)
+        self.assertEqual(set(words), set(self.interesting_alert_word_list) - {'alert'})
+
     def test_remove_word(self) -> None:
         """
         Removing alert words works via remove_user_alert_words, even