Update username checks

[ChrisLovering]

Closes #1891

Currently, we check people's usernames when they send a message on the server, to ensure it isn't inappropriate. To get more coverage, we should actually be checking member join and member update events, as people can still have their nickname be visible on server without ever sending a message. They can also update their usernames after sending a message.

I have also:

• Added a check for the normalised version of the name too, using unicodedata.normalize, to ensure unicode versions of names don't bypass our filters.
• Returned the first bad match we find, rather than collecting them all, so we don't run it against all of our filters if not needed. If a single bad match is found, I trust our mods to actually read the user's name
• Moved the check_send_alert() call to the first thing we do, as a redis call is cheaper than checking all our filter tokens

[HassanAbouelela]

All clear! The alphabetical sorting was unexpected, but appreciated, nicely done.

[wookie184]

Normalizing strings before checking if they match on all filters removes our ability to match specific unnormalized forms. I guess this should be ok, although we should look into our current filters and see what can no longer match. I'm not sure how aggressive the normalization by confusables is.

[wookie184]

I had a bit more of a look into the string normalisation stuff, so I just want to outline my findings.

1. Just as a note, we should definitely not use confusables.normalize on the whole string, it has... problems. For some letters it seems to have multiple possible confusable characters, and instead of choosing one, it returns a list of all the possible combinations. If somebody sent 100 Ｓ characters to the bot (it seems to think it could be an f or an s 🤷), it would have tried to generate a list of length 2**100 (1267650600228229401496703205376).
2. unicodedata.normalize seems pretty good, although it can't remove accents entirely by itself. However, we can remove them by using the NFKD form which decomposes accented letters into the letter and accent characters separately. We can then remove the accent by filtering whether it is a combining character with unicodedata.combining

''.join(
        c
        for c in unicodedata.normalize("NFKD", text)
        if not unicodedata.combining(c)
    )

3. That still doesn't normalize quite a lot of other letters, e.g. cyrillic letters. If we did want to match that, we could use confusables.normalize as well but on individual characters at a time to avoid the problem mentioned above (although of course that still wouldn't be perfect).

Our solution doesn't need to be perfect so we can probably just go with the simple option (2) and revisit it if we feel it's necessary.

[onerandomusername]

this looks good to me, keep up the moderation tooling!

(side: I agree with bluenix's comments)

[jchristgit]

Nice use of unicodedata 👍