feat: harden HTTP resiliency with connection retries, request timeout, and docs

[iMicknl]

Summary

Fixes #2147.

Hardens the client's behaviour on transient network failures and documents the full resiliency strategy. Builds on the original connection-retry centralization with a request timeout, smarter ServerDisconnectedError handling, and user docs.

Changes

Centralize connection retry (original fix)

Previously only fetch_events was decorated with @retry_on_connection_failure, so a transient TimeoutError / aiohttp.ClientConnectorError raised from any other method — including the command path (execute_action_group -> _execute_action_group_direct) and all setup/state/refresh calls — propagated raw on the very first occurrence instead of being retried.

This surfaced in Home Assistant (home-assistant/core#173155): a ConnectionTimeoutError (subclass of the builtin TimeoutError) raised from a cover close command escaped as an unhandled traceback.

The retry is now centralized on the _get/_post/_put/_delete request helpers so every request gets uniform transient-connection retry, with connection retry sitting innermost (closest to the request). The budget is intentionally fail-fast: 3 tries within ~30s.

Default request timeout

The client-owned session now applies a default per-request timeout (total=15s, sock_connect=10s), kept below the connection-retry budget so a hung socket fails fast as a TimeoutError and is retried, instead of blocking on aiohttp's 300s default. max_time only bounds the scheduling of retries between attempts; it cannot interrupt an in-flight request, so the timeout is what makes the fail-fast budget meaningful at runtime.

Callers passing their own ClientSession own its timeout; pyOverkiz does not override it.

ServerDisconnectedError no longer forces a relogin

ServerDisconnectedError (a dropped keep-alive socket) was bucketed with NotAuthenticatedError and triggered a full login() + event-listener re-register on every occurrence. It is now treated as a transient transport failure and simply retried, with no relogin. A genuine session expiry is still reported by the server as a Not authenticated response, which raises NotAuthenticatedError and escalates to relogin via the separate auth policy.

Docs

Adds docs/resiliency.md explaining what the library retries (with budgets and on-retry actions), the fail-fast rationale, exponential backoff with full jitter, request-timeout behaviour and how to configure a custom session, and which exceptions consumers should handle themselves. Linked from the nav and cross-referenced from troubleshooting.md.

Tests

• test_backoff_retries_command_on_connection_failure / test_backoff_gives_up_after_max_tries_on_connection_failure — command/GET path connection retry and give-up.
• test_backoff_retries_on_server_disconnected_without_relogin — a disconnect is retried without calling login().
• test_server_disconnected_escalates_to_relogin_when_auth_expired — a retry that surfaces a real auth expiry still triggers relogin.
• test_default_session_has_request_timeout — the client-owned session applies a default timeout within the retry budget.

Verification

• 552 tests pass
• ruff clean
• mypy clean on client.py