Add more flexibility to configure LLM model parameters

[p-]

Currently the LLM parameters can only be configured directly in the code (with exception of some ENV variables like MODEL_TEMP (temperature)).
See model settings for supported values. Discuss which settings should be included (e.g. reasoning settings for GPT-5) or if it should be possible to pass in a dict.

[m-y-mo]

Maybe can pass it as a dict in model_config and then apply a filter to remove keys that aren't supported before passing it to ModelSettings here? e.g.

async def deploy_task_agents(available_tools: AvailableTools,
                            ...
                            model_params: dict,...): 
...
   model_params['tool_choice'] = ('auto' if toolboxes else None)
   filter_unsupported_keys(model_params)
   model_settings = ModelSettings(**model_params)

something like that and create model_params from model_config of the task or use default?