Add model and provider parameters to delegate_task tool schema and
handler, allowing the LLM to specify which model/provider to use for
each delegation call (overriding delegation config).

Also fix credential resolution: when delegation.base_url is set but no
api_key is found AND a provider is configured, fall through to provider
resolution instead of raising an error. This handles the common case
where base_url is redundant but provider can resolve creds from env.

…l selection

Map task categories (coding, research, reasoning, etc.) to specific models
via delegation.task_models in config. The agent passes task_type when
delegating, and the model is resolved automatically.

Resolution order: per-call model > per-task model > task_models[type] >
delegation.model > parent model. task_models only selects the model name;
provider is always delegation.provider (or per-call provider, or parent).

Also adds per-task model override in batch mode, task_type param in schema
with documentation hints, and user-facing docs with examples.