feat: route GET/HEAD requests to an optional read replica - #71

Closes #30.

What

• New DB_REPLICA_DSN setting (default empty). When set, GET/HEAD requests use a replica engine; every other method, and any resolution without a request, uses the primary. When unset, behaviour is unchanged and no second pool is created.
• app/ioc.py: database_engine (primary), database_replica_engine (AsyncEngine | None), and a request-scoped dynamic_engine chosen by choose_sa_engine. The session gets dynamic_engine via explicit kwargs.
• Tests override dynamic_engine with the rolled-back connection, so reads and writes still share one transaction.
• tests/test_db_routing.py covers method routing, no-request fallback, no-replica fallback, and building the replica engine from the DSN.

Design notes

• Modelled on the rchat chats service, with two deliberate differences:
  • The session is wired to dynamic_engine explicitly instead of relying on bound_type=None making it the only AsyncEngine for type autowiring. Both base engines are bound_type=None so modern-di does not raise a duplicate-type error.
  • A separate DB_REPLICA_DSN instead of a db-retry multi-host DSN with target_session_attrs. It keeps the template dependency-free and the config obvious; db-retry also requires a separate database name.
• The replica provider returns None rather than aliasing the primary provider, so the routing is testable by overriding each engine independently.
• HEAD goes to the replica alongside GET, since it is equally read-only.
• app/resources/*.py is exempted from ruff's TC rules: modern-di evaluates creator annotations at runtime, and ruff's autofix moving import fastapi under TYPE_CHECKING broke resolution with NameError.

Caveats

• Replica lag: a client that writes and immediately reads may see stale data.
• GET handlers must not write; they would hit the replica.