Handlig unicode on annotation creation

This is an odd bug! On specific browser, they have noticed that on Label
Studio this gives the following issues:

```
psycopg2.errors.InvalidTextRepresentation: invalid input syntax for
type json
DETAIL: Unicode low surrogate must follow a high surrogate
```

We work around this by doing a round trip to JSON, with
`ensure_ascii=False` will ensure that any UNICODE characters are
preserved, and therefore escaped on the round trip back.

Note: we don't care about any unicode in annotation, as we don't use
the text for referencing our annotation.

Test Plan:

We can reproduce this by forcing the annotation end point to consume unicode.

1. Use the `copy as CURL` option on your browser to copy the "update annotation" call from Label Studio

2. Inject some unicode into the CURL command - e.g.

```
curl 'http://localhost:8080/api/annotations/126?taskID=5979&project=43' \
  -X 'PATCH' \
  #.... snipped
  --data-raw $'{"result":[{"value":{"value":{"start":"/div[1]/div[1]/text()[1]","startOffset":0,"end":"/div[1]/div[1]/text()[1]","endOffset":88,"globalOffsets":{"start":0,"end":88},"text":"Some unicode: \udfff	"  ...],"draft_id":0,"parent_prediction":null,"parent_annotation":null,"project":"43"}'
```

3. Run it againt the dev server, previously this will cause 500, and with the fix it should 200.

label_studio/tasks/serializers.py

@@ -75,6 +75,36 @@ def create(self, *args, **kwargs):
 
     def validate_result(self, value):
         data = value
+
+        # This is to work around a mysterious bug related to DRF handling:
+        # https://github.com/encode/django-rest-framework/issues/7026
+        #
+        # This manifest itself as 500 error when writing to database:
+        #
+        #   psycopg2.errors.InvalidTextRepresentation: invalid input syntax for
+        #   type json
+        #   DETAIL: Unicode low surrogate must follow a high surrogate
+        #
+        # This only happens to specific browser, presumably some bad unicode
+        # was leaked to the server.
+        #
+        # We work around this by doing a round trip to JSON, with
+        # ensure_ascii=False will ensure that any UNICODE characters are
+        # preserved, and therefore escaped on the round trip back.
+        #
+        # Note: we don't care about any unicode in annotation, as we don't use
+        # the text for referencing our annotation.
+        if not isinstance(value, str):
+            json_str = json.dumps(value, ensure_ascii=False)
+            # Strip unicode characters
+            non_unicode_json_str = json_str.encode('ascii', 'ignore').decode('ascii')
+
+            # If the string is different, then we have some unicode characters
+            # then we will do the roundtrip de-unicoding.
+            if json_str != non_unicode_json_str:
+                logging.warning('Annotation result contains unicode characters, stripping them')
+                value = non_unicode_json_str
+
         # convert from str to json if need
         if isinstance(value, str):
             try:
@@ -282,7 +312,7 @@ def _insert_valid_user_reviews(dicts, members_email_to_id, default_user):
             if email not in members_email_to_id:
                 obj['created_by_id'] = default_user.id
                 logger.warning('Email not found in members_email_to_id, default user used instead')
-                
+
             # resolve annotators by email
             else:
                 obj['created_by_id'] = members_email_to_id[email]
@@ -427,7 +457,7 @@ def add_reviews(self, task_reviews, annotation_mapping, project):
         """ Save task reviews to DB
         """
         return []
-    
+
     def add_drafts(self, task_drafts, db_tasks, annotation_mapping, project):
         """ Save task drafts to DB
         """