fix: name psycopg2 in the Alembic DSN and cap sqlalchemy below 2.1

[lesnik512]

Closes #38. Follow-up to lift the cap: #40.

What

• Settings.sync_db_dsn_parsed sets drivername="postgresql+psycopg2" instead of "postgresql"; the test that pinned the old value now pins the driver.
• sqlalchemy[asyncio]<2.1.

Why

SQLAlchemy 2.1.1 (2026-09-25) arrived through the fresh resolve and broke the pytest jobs in three independent ways:

1. postgresql:// now means psycopg (3), so alembic upgrade head imported a driver this project does not depend on. Naming the driver fixes the URL's meaning regardless of SQLAlchemy's default (option 1 from CI: pytest jobs fail on a fresh resolve because SQLAlchemy 2.1 resolves postgresql:// to psycopg 3 #38; psycopg 3 is a separate decision).
2. advanced-alchemy 1.11.0 (latest) does not translate 2.1's asyncpg integrity errors: its regexes expect a DETAIL line that 2.1 no longer includes (Bug: DuplicateKeyError not raised on SQLAlchemy 2.1 with asyncpg (regex no longer matches) litestar-org/advanced-alchemy#803, fix PR open). Duplicate username and unknown member id return 500 instead of 409/400. Confirmed independent of anything here by running the suite on 2.1.1 with this repo's models untouched.
3. Two deprecations surface as errors under filterwarnings = ["error"]: advanced-alchemy's @declarative_mixin (#804) and this repo's lazy="noload".

(2) has no fix on our side, so the cap is the honest change; (3) is recorded in #40 with the verified replacement (lazy="raise"), to land when the cap lifts.

Verified

• Red first: test_sync_db_dsn_parsed_names_the_sync_driver fails against the old drivername ('postgresql' == 'postgresql+psycopg2'), green after the settings change.
• Full suite in Docker on the capped resolve (SQLAlchemy 2.0.54): 117 passed, 100% coverage.
• ruff format, ruff check, ty clean.