fix: Services survive config entry reload, get_history returns a dict response

- service registration extracted to idempotent _async_register_services,
  called from both async_setup and async_setup_entry: unloading the last
  entry unregisters services, and a reload (every options change) never
  re-ran async_setup, leaving the integration without services
- get_history handler wraps the list in {"history": [...]}: HA rejects
  non-dict action responses, so every return_response call failed with a
  server error since the service gained SupportsResponse (v2.4.x) - the
  path never worked, no consumer could depend on the old shape
- README: get_history example shows response_variable usage and the
  actual limit clamp semantics

Both found by live smoke test in HA 2026.7 (Docker), not by static review

README.md

@@ -283,12 +283,13 @@ data:
 ```yaml
 service: ha_text_ai.get_history
 data:
-  limit: 5  # optional, number of conversations to return (1-100)
+  limit: 5  # optional, number of conversations to return (values above 200 are clamped); omit to get the full stored history
   filter_model: "gpt-4o"  # optional, filter by specific AI model
   start_date: "2025-02-01"  # optional, filter conversations from this date
   include_metadata: false  # optional, include tokens, response time, etc.
   sort_order: "newest"  # optional, sort order: "newest" or "oldest"
   instance: sensor.ha_text_ai_gpt
+response_variable: history_result  # entries are in history_result.history
 ```
 
 ## 🚀 Advanced Automation Examples with Response Variables