fix: make is_dsn_multihost agree with the connection plan

[lesnik512]

Closes #36.

is_dsn_multihost and build_connection_plan each decided what "multi-host" means. They used different checks, and they disagreed:

DSN query	is_dsn_multihost before	after	build_connection_plan
host=a:5432&host=b:5433	True	True	fails over to 2 hosts
host=a,b&port=5432,5433	False	True	fails over to 2 hosts
host=a&host=b (no ports)	True	raises ArgumentError	raises ArgumentError
host=a, a:5432/db, no host	False	False	no failover

Now there is one definition. parse_connect_args(url) runs the asyncpg dialect's parse once and returns the connect args together with the (host, port) pairs to fail over to. build_connection_plan and is_dsn_multihost both use it. is_dsn_multihost is True exactly when the plan would have failover.

The PR also stops documenting the portless multi-host DSN, which the connection factory rejects. That shape was in the build_db_dsn docstring and the test fixtures. They now use host=h1:p1&host=h2:p2.

Release note

Behaviour change in is_dsn_multihost:

• It now returns True for the comma-separated format, ?host=h1,h2&port=p1,p2.
• It now raises sqlalchemy.exc.ArgumentError for DSNs the connection factory would reject, for example repeated host params without ports. It used to return True for these.

Testing

• New tests fail on main: a parametrised check that is_dsn_multihost agrees with build_connection_plan for each DSN shape, the comma format, and the ArgumentError case.
• I ran the full suite against Postgres 18 locally: 45 passed, 100% coverage. ruff, ty and eof-fixer are clean.