Allow each task in the 'tasks' batch array to specify its own model
and provider, routing individual subtasks to different LLMs while the
rest inherit the global delegation config.

Problem: delegate_task applied a single global model (from
delegation.model in config.yaml) to every child. There was no way to
say 'run this hard reasoning task on Opus, everything else on Sonnet'
without changing the global config.

Solution: add optional model and provider fields to each task object
in the tasks array. When present, _build_child_agent is called with
a task-specific credential bundle resolved from a config overlay
(global config + per-task overrides). When absent, the task falls back
to the existing global creds — no behavior change for existing callers.

Changes:
- tools/delegate_tool.py: resolve per-task creds when task.model or
  task.provider is set; pass task_creds into _build_child_agent
  instead of the global creds for that task
- tools/delegate_tool.py: add model and provider to DELEGATE_TASK_SCHEMA
  tasks[].items.properties with descriptive docstrings
- tools/delegate_tool.py: add MODEL ROUTING section to tool description
- tests/tools/test_delegate.py: TestPerTaskModelOverride (4 tests)

Note: cron jobs already support per-job model/provider (cronjob_tools.py).
This brings delegate_task to parity. Both use OpenRouter's model routing
under the hood — provider='openrouter' + any model string works today.