fix: clear rate-limit hot-settings cache and add retry loop for maintenance-mode tests

- Invalidate `_hot_settings_value` and `_hot_settings_at` in `_patch_limits` so stale
  cached limits from prior tests do not persist after monkey-patching `get_int_setting`.
- Add a polling retry loop (up to 5 seconds) in `test_maintenance_mode_blocks_guests`
  to handle eventual consistency of the maintenance-mode flag propagation.
- Replace brittle hardcoded `@bob_test` assertion in mention test with dynamic
  `data-username` attribute extraction from the active dropdown item.
- Replace `asyncio.run` calls in bot unit tests with shared `run_async` helper from
  `tests.conftest` to avoid event-loop conflicts in the test suite.
- Change `insert` to `upsert` with `["uid"]` conflict target in Gitea test fixture
  to prevent duplicate-key errors on repeated test runs.
- Replace `asyncio.new_event_loop().run_until_complete` in Telegram worker tests
  with the same shared `run_async` helper for consistency.

tests/e2e/post.py

@@ -158,8 +158,11 @@ def test_mention_arrow_then_enter_inserts(alice):
         state="visible", timeout=5000
     )
     textarea.press("ArrowDown")
+    selected = page.locator(".mention-dropdown-item.active").first
+    selected.wait_for(state="visible", timeout=5000)
+    username = selected.get_attribute("data-username")
     textarea.press("Enter")
-    expect(textarea).to_have_value(re.compile(r"@bob_test\s"))
+    expect(textarea).to_have_value(re.compile(rf"@{re.escape(username)}\s"))
 
 
 def test_mention_escape_closes_dropdown(alice):