fix(cli): fail check on database login errors and read ~/.pgpass

[psteinroe]

check no longer passes silently when the configured database can't be used, and it reads the password from ~/.pgpass / PGPASSFILE the way libpq clients do.

Connection failures fail check

When a database is configured (not disabled), check connects and loads the schema before checking the files. If that fails, it reports a database/connection error, still runs the checks that don't need the database, and exits non-zero. This includes a connection that is still in backoff from an earlier failure in the daemon. The error is visible without a log level:

database/connection ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

  × Database error: password authentication failed for user "postgres"

  i The checks that need the database were skipped. Fix the connection settings, or pass --disable-db or set db.disableConnection to check without a database.

This adds a check_database_connection workspace method, which the daemon also serves. The connection outcome is returned as a result value, not an error, because workspace errors lose their type over JSON-RPC. The LSP doesn't call the new method, so its behaviour is unchanged: it backs off and keeps working without the database.

The DatabaseConnectionError diagnostic template didn't interpolate, so dblint printed the literal Database error: {message}. That's fixed too.

disableConnection in the config file is honoured by the CLI

--disable-db was declared as switch with fallback(Some(false)), so the CLI's partial configuration always carried disableConnection: false and overrode true from the config file. That was invisible while failures were silent; it would now fail check. The flag now only sets the value when passed.

Password file

With no password from the config, --password, or PGPASSWORD, the password is looked up in PGPASSFILE, else ~/.pgpass (%APPDATA%\postgresql\pgpass.conf on Windows). If nothing matches, it falls back to the documented default postgres. That fallback is the conservative choice: setups that rely on it keep working.

sqlx 0.8 has pgpass support, but it isn't usable here:

• It runs only inside PgConnectOptions::new() / URL parsing. We called .host().port().username() after that, so it matched against the environment's defaults, and then we always overwrote the password.
• It's crate-private and doesn't expose whether it found a password, which we need for the fallback.
• It deviates from libpq: it doesn't unescape \: / \\ in the password, it falls back to ~/.pgpass when PGPASSFILE has no match, and it checks permissions only on Linux.

So pgpass.rs ports libpq's passwordFromFile/pwdfMatchesString (REL_18_6). Like libpq, it ignores the file on Unix if group or others can access it, or if it isn't a regular file. Connection strings without a password use the same lookup, and keep sqlx's result when it finds nothing.

Follow-ups, not changed here

• --conn_timeout_secs has the same fallback(Some(10)) pattern, so the CLI overrides connTimeoutSecs from the config file.
• In daemon mode (--use-server), file diagnostics fail with "Method not found". The client requests pgls/pull_diagnostics, but the server registers pgls/pull_file_diagnostics. This is already the case on main.
• For an unreachable server, sqlx reports pool timed out while waiting for an open connection rather than the underlying error.

Fixes #834