fix: prune stale in-memory rate limit keys

[saurabhhhcodes]

Summary\n\nCloses #928\n\nThe shared distributed rate limiter now periodically prunes stale in-memory keys instead of letting the fallback dict grow forever. It also removes the direct _scan_attempts_by_user alias from the Guard router and moves tests onto public helper methods so the limiter internals stay encapsulated.\n\n## Changes\n\n- add public clear_local_attempts() and cleanup_stale_local_attempts() helpers\n- track the longest window seen per key so stale cleanup stays safe across endpoints\n- prune expired local entries during periodic cleanup sweeps\n- remove the direct internal alias from backend/app/api/v1/guard.py\n- update conftest and Guard tests to use the public helper instead of private state\n\n## Validation\n\n- git diff --check\n- /Users/saurabhkumarbajpaiai/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/bin/python3 -m pytest backend/tests/test_rate_limit_helper.py backend/tests/test_guard_api.py backend/tests/integration/test_rate_limiting.py -q\n\n## Notes\n\nThe broader backend suite still has unrelated deprecation warnings in the current checkout, but the touched rate-limit paths passed cleanly.