chore: drop the asyncpg OpenTelemetry instrumentor

[lesnik512]

Summary

The app ran two database instrumentors, so every query was traced twice. This keeps SQLAlchemyInstrumentor and removes AsyncPGInstrumentor(capture_parameters=True) and its package.

With one instrumentor left, nothing else would notice if it stopped producing spans. A new test, tests/test_tracing.py, turns tracing on, makes a request, and checks that the SQLAlchemy instrumentor records a SELECT span under the request span.

Why

Measured against Postgres 17 in Docker on the same host, with a new session per request and 100 concurrent workers:

instrumentation	requests/s	spans per request
none	~4.9k	0
SQLAlchemy + asyncpg	~2.4k	8
SQLAlchemy only	~3.5k	2

The extra asyncpg spans were BEGIN, ROLLBACK and the pool pre-ping queries. capture_parameters=True also copied query parameter values, user data included, into span attributes.

What we lose

• COMMIT no longer gets its own span. Its time still shows inside the request span.
• Failed queries keep ERROR status and the error message, but there is no exception event with a traceback.

Watch out

opentelemetry-instrumentation-sqlalchemy (0.65b0, and 0.66b1 too) declares sqlalchemy < 2.1. On SQLAlchemy 2.1 it skips instrumentation, and lite-bootstrap silences the log line that reports the conflict. Keep the sqlalchemy<2.1 pin until a release ships 2.1 support (open-telemetry/opentelemetry-python-contrib#5120). The new test fails in that case.

Test plan

• just lint and the full test suite pass, with coverage at 100%.
• With SQLAlchemyInstrumentor removed from build_app, the new test fails.