Implement ai.workers: Worker, WorkerProvider and Session

One of two competing drafts for ai.workers. The other is #347.

A worker is one session of a coding-agent CLI (Codex, Claude Code, Pi). In this draft it has the shape of an ai.Model: plain data and a provider.

• Worker: the settings and the state (session_id, messages).
• WorkerProvider: name, executable, env, and the stream() that ai.stream and ai.Agent.run call.
• Session: direct control (start, send, events, interrupt, close).

worker = workers.Codex(cwd=".")
async with worker.session() as session:
    async with agent.run(worker, messages) as run: ...
    await session.send(ai.user_message("And the tests?"))

Compared with #347

• Core changes less. ai.stream still calls model.provider.stream(model, ...); the diff outside ai/workers is Model -> ModelLike annotations, one aclosing block in api.py, the Context.model annotation, and a new models/core/model_like.py with the two protocols. model.py, hooks.py and ai/__init__.py are as on main.
• Direct control is on a separate Session, not on the worker. The CLI starts on first use inside async with worker.session(), and a model call outside that block raises.
• Not here, and in #347: approvals through hooks, output_type, fork() and sync().

Status

• Offline tests, ruff, mypy (src tests) and ty pass locally.
• The live tests (AI_WORKERS_LIVE=1 uv run pytest tests/workers/test_live.py) were not run on the last commit.