Switch event classes to dataclasses

[pgjones]

The _EventBundle used previously dynamically creates the Event
classes, which makes typing hard. As the Event classes are now stable
the dynamic creation isn't required, and as h11 supports Python3.6+
dataclasses can be used instead.

This now makes the Events frozen (and almost imutable) which better
matches the intended API. In turn it requires the object.__setattr__
usage and the alteration to the
_clean_up_response_headers_for_sending method.

This change also improves the performance a little, using the
benchmark,

Before: 6.7k requests/sec
After: 6.9k requests/sec

Notes:

The test response-header changes are required as the previous version
would mutate the response object.

The init for the Data event is required as slots and defaults aren't
possible with dataclasses.

[bluetech]

I am personally not a big fun of using dataclasses or attrs in a public library API, because

1. They add some startup overhead (probably negligible in this case though).
2. They add "hidden" APIs to the classes, like dataclasses.asdict, dataclasses.replace etc. which users will eventually rely on, making it part of the public API.
3. They somewhat reduce (or add friction to) the library's control should it ever need it.
4. They tend to do some funky stuff, for example in pytest we had some trouble with what dataclasses does here.

That said, if you're OK with this, the code LGTM except one comment.