[fix] run_response.model: update from actual provider response model name after LLM call

[NIK-TIGER-BILL]

Problem

run_response.model is set to agent.model.id before the LLM call and never updated afterwards (closes #6921).

When a LiteLLM router/proxy switches models via a fallback (e.g. DeepSeek → Qwen), the API response object's model field contains the actual model name — but it was silently ignored. Callers who read run_output.model to identify the model that was used see the originally configured model name, not the one that actually served the request.

Fix

1. In OpenAIChat._parse_provider_response(), save response.model into model_response.provider_data['model'] when it is present:

if getattr(response, 'model', None):
    model_response.provider_data['model'] = response.model

2. In update_run_response(), update run_response.model from provider_data['model'] if available:

_actual_model = model_response.provider_data.get('model')
if _actual_model:
    run_response.model = _actual_model

Closes #6921