Fix thread-unsafe state writes; use conversation-agent dropdown

- entity.py: decorate the dispatcher target _handle_update with @callback so it
  runs on the event loop. Without it, async_dispatcher_send offloaded the plain
  sync callback to an executor thread, and async_write_ha_state raised
  "calls async_write_ha_state from a thread other than the event loop".
- config_flow.py: replace the free-text Assist agent field with
  ConversationAgentSelector (a proper agent dropdown), and use LanguageSelector
  for the language field.
- README: clarify that the built-in Assist agent only handles device commands;
  an LLM conversation agent is required for free-form chat.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -95,8 +95,12 @@ will answer.
 ### Options (⚙ → Configure)
 
 - **Route incoming messages to Assist** – turn the chat bridge on/off.
-- **Assist agent / language** – pick a specific conversation agent, or leave
-  blank for the default pipeline.
+- **Assist agent / language** – pick the conversation agent from the dropdown.
+  > **Want free-form chat?** The default *Home Assistant* agent only recognises
+  > device-control commands and will answer anything else with "Sorry, I
+  > couldn't understand that". To actually chat, install an LLM conversation
+  > integration (OpenAI, Google Generative AI, Anthropic, Ollama, …), then
+  > select that agent here.
 - **Reply to any sender** / **Allowed sender addresses** – restrict who Assist
   will answer.
 - **Default recipient** – address used by the `notify.reticulum` entity.