tornado: Replace dataclasses.asdict() call, as it is slow.

This code is called in the hot path when Tornado is processing events.
As such, making this code performant is important.  Profiling shows
that a significant portion of the time is spent calling asdict() to
serialize the UserMessageNotificationsData dataclass.  In this case
`asdict` does several steps which we do not need, such as attempting
to recurse into its fields, and deepcopy'ing the values of the fields.

In our use case, these add a notable amount of overhead:
```py3
from zerver.tornado.event_queue import UserMessageNotificationsData
from dataclasses import asdict
from timeit import timeit
o = UserMessageNotificationsData(1, False, False, False, False, False, False, False, False, False, False, False)
%timeit asdict(o)
%timeit {**vars(o)}
```

Replace the `asdict` call with a direct access of the fields.  We
perform a shallow copy because we do need to modify the resulting
fields.

zerver/tornado/event_queue.py

@@ -9,7 +9,6 @@ import traceback
 import uuid
 from collections import deque
 from contextlib import suppress
-from dataclasses import asdict
 from functools import lru_cache
 from typing import (
     AbstractSet,
@@ -987,7 +986,11 @@ def process_message_event(
             all_bot_user_ids=all_bot_user_ids,
         )
 
-        internal_data = asdict(user_notifications_data)
+        # Calling asdict would be slow, as it does a deep copy; pull
+        # the attributes out directly and perform a shallow copy, as
+        # we do intend to adjust the dict.
+        internal_data = {**vars(user_notifications_data)}
+
         # Remove fields sent through other pipes to save some space.
         internal_data.pop("user_id")
         internal_data["mentioned_user_group_id"] = mentioned_user_group_id