refactor(gui-personalization): extract endpoint to a module + fix model defaults

[ArmenSl]

Cleanup on top of #577, addressing the organization + correctness review. Targets feature/gui_personalization so it lands before #577 merges to development.

Organization (the main concern)

The new /personalize-gui-page endpoint inlined ~150 lines into generation_router.py — validation, a ~40-line LLM system prompt, the OpenAI call and response parsing — which breaks the codebase's own pattern (agent_personalization.py holds that logic; the router just calls it).

• New services/utils/gui_personalization_utils.py — mirrors agent_personalization: the system prompt is a module constant, and personalize_gui_page() owns validation → prompt → LLM call → parse.
• The router endpoint is now thin — resolve API key, asyncio.to_thread(run_gui_personalization, …), shape the response; @handle_endpoint_errors maps ValidationError/GenerationError.

Correctness

• Model defaults were inconsistent and invalid — gpt-5.4-mini (×7) and gpt-5.6-luna, neither a real OpenAI id (would 404 at runtime). Consolidated to a single DEFAULT_OPENAI_MODEL = "gpt-5" constant, used everywhere.
• Removed a leftover print("calling now") debug statement.
• Removed two dead … and False branches in replace_reply_batch (a disabled translate_text_api experiment).

Tests

• test_gui_personalization.py — 10 tests (LLM mocked, offline): input validation, default-vs-explicit model, response parsing, and error mapping (missing key → ValidationError, unparseable/invalid response → GenerationError).
• Full converter round-trips + agent-personalization + agent-converter suites still pass (94 tests).

Deliberately left

• The translate-before-style reorder in replace_reply_batch — appears intentional; flag for @Aran30 to confirm.
• Pre-existing gpt-5.5 default at generation_router.py:172 (a different, older endpoint) — out of this PR's scope.
• The now-unused translate_text_api helper is left in place (removing it is separate).