feat: support OAUTHBEARER SASL authentication - #1378

PostgreSQL 18 added the `oauth` HBA method, which authenticates a
connection with an OAuth 2.0 bearer token over the SASL OAUTHBEARER
mechanism (RFC 7628) instead of a password. `authenticate_sasl`
recognised only the SCRAM mechanisms, so such a server was refused with
"unsupported SASL mechanism".

The caller supplies the token through `Config::token_provider`, a closure
called once per connection attempt so that an expiring token can be
refreshed without rebuilding the config. Obtaining, caching and renewing
tokens stays out of scope, as it is in libpq when a custom flow is
installed: there is no discovery and no device authorization flow. No
token is requested unless the server asks for OAUTHBEARER.

A successful exchange is a single round trip. The bearer token travels in
the SASL initial client response and the server replies with
`AuthenticationOk`; OAUTHBEARER carries no additional data on success, so
there is no `AuthenticationSASLFinal` message. On failure the server sends
a challenge holding a JSON error document, which RFC 7628 Sec. 3.2.3
requires the client to answer with a lone kvsep byte - sending it is what
gets the server to report the failure as its own `ErrorResponse` instead
of the client giving up with an unexpected-message error.

OAUTHBEARER defines no channel binding and the server rejects the `p`
specifier, so the gs2 header is always `n,,` and `channel_binding=require`
is refused.

Picks up the approach from rust-postgres#1325 by @JoHaHu.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

`cargo clippy --all --all-targets` fails on master with `RUSTFLAGS=-Dwarnings`:

    error: this lifetime isn't used in the type
      --> tokio-postgres/tests/test/types/mod.rs:51:24
       |
    51 |     T: PartialEq + for<'a> FromSqlOwned + ToSql + Sync,
       |                        ^^

`FromSqlOwned` is declared as `pub trait FromSqlOwned: for<'a> FromSql<'a> {}`,
so it already carries the higher-ranked bound internally and takes no lifetime
parameter of its own. `for<'a> FromSqlOwned` therefore binds a lifetime nothing
uses, and removing the binder is behaviour-preserving by construction rather than
by inspection.

Not a new defect. `extra_unused_lifetimes` learned to see through an HRTB in a
clippy released after master's last CI run, and the workflow pins
`dtolnay/rust-toolchain@stable` rather than a version, so the next push to touch
CI meets the newer lint. Any PR opened now is red before it changes anything.

Reproduced with CI's own invocation and toolchain, clippy 0.1.98:

    RUSTFLAGS=-Dwarnings cargo clippy --all --all-targets
      without this change: exit=101
      with this change:    exit=0

`cargo fmt --all -- --check` clean, and `cargo test -p tokio-postgres --test
test --no-run` still builds the target.