add validation for isupport tokens

According to jwheare, ISO 8601 doesn't specify the behavior one way
or the other, and the server-time spec is also ambiguous, but this
is safest.

Fix spurious bidi rule violations in casefolding channel names
by stripping the # before starting the casefolding.

* fix a race condition: a call to `Write` does not spawn a writer goroutine
  if the trylock is held, so `BlockingWrite` must check for fresh data after
  releasing the trylock
* streamline some close/finalize logic

1. Handle PROXY lines with IPv6 addresses starting with ::
(similar to WEBIRC in issue #211)

2. Strip v6 mapping from v4 addresses when handling proxied IPs.

According to https://defs.ircdocs.horse/defs/numerics.html, 330 RPL_WHOISACCOUNT
takes 4 parameters: `<client> <nick> <authname> :<info>`. We were omitting
the second parameter (the target nick).

Also refactor locking.

The server should never crash during rehash, even if the config is invalid.

Previously, we generated and prepended a long salt before generating
password hashes. This resulted in the hash verification cutting off long
before it should do. This form of salting is also not necessary with
bcrypt as it's provided by the password hashing and verification
functions themselves, so totally rip it out.

This commit also adds the functionality for the server to automagically
upgrade users to use the new hashing system, which means better
security and more assurance that people can't bruteforce passwords.

No need to apply a database upgrade to do this, whoo! \o/