feat: ship PrometheusRedisMetricsProvider and a cached get_redis_metrics

[AlexeyShalaev]

Closes #36.

Problem

AsyncRedisProvider reads RedisMetricsProtocol | None and RedisMetrics implements it, but the kit ships nothing that provides the one from the other, so every consumer carries the same dishka provider. And RedisMetrics(prefix=None) twice raises Prometheus's duplicated-timeseries error, so a container rebuilt per test cannot ask for the collector twice.

What ships

Two feat commits.

redis_client_kit.metrics.get_redis_metrics(prefix=None) — one RedisMetrics per prefix on the default registry, cached, the shape of grpc-client-kit's get_grpc_client_metrics. "" and None are the same unprefixed instance, because RedisMetrics treats them alike. RedisMetrics itself is unchanged: building two by hand still raises, and the docstring and the docs say that is Prometheus and point at the getter.

redis_client_kit.providers.PrometheusRedisMetricsProvider(*, prefix=None) — Scope.APP, requests RedisSettingsProtocol, provides RedisMetricsProtocol | None as get_redis_metrics(prefix) when settings.metrics_enabled and None otherwise, so it drops into AsyncRedisProvider(provide_default_metrics=False) (or after AsyncRedisProvider()) and the client is instrumented exactly when the settings say so. RedisSettingsProtocol gains metrics_enabled: bool for that; BaseRedisSettings has carried the field from the start, a hand-written settings class needs the one attribute, and the configuration guide's example now has it. With metrics_enabled on and the metrics extra missing, resolving the collector raises the ImportError naming the extra, the way every other extra here fails.

Rejected

• The reporter's PrometheusMetricsSettingsProtocol + default_prefix keys (sqlalchemy-foundation-kit's shape). Those are the application's keys, and a kit provider that requests them makes every consumer provide them. The prefix is a constructor argument, as in grpc-client-kit.
• A warning and None when the extra is missing (grpc-client-kit's choice). Silent in production; this kit fails loud on a missing extra everywhere else.
• A registry= parameter on RedisMetrics. Nothing here needs it.

Tests

• Getter: the same prefix twice is the same object; different prefixes differ; "" is None; RedisMetrics twice raises.
• Provider, on the reporter's container (AsyncRedisProvider(provide_default_metrics=False), PrometheusRedisMetricsProvider()): metrics off gives None and a plain Redis; metrics on gives InstrumentedRedis whose _metrics is get_redis_metrics(); prefix= names the series; two containers in a row hand out the same collector; registered after AsyncRedisProvider() it wins; a missing extra raises ImportError naming it.
• test_metrics_init.py plants its mocked modules through monkeypatch now, so they leave sys.modules after each test instead of staying there for the rest of the session (they fed the provider's lazy import a MagicMock).

Gate: uv run ruff check ., uv run ruff format --check ., uv run mypy redis_client_kit, uv run pytest -m unit — 160 passed, unit coverage 99.66%.

Docs: advanced guide (metrics and Dishka sections), configuration guide, agents page, README.