Add cron_effort to run cron/hook turns below interactive effort

[alex-fedotyev]

Summary

Cron- and hook-sourced turns are sensing and triage work that fire far more
often than interactive sessions and rarely need the interactive model's full
reasoning depth. Today they inherit the global agent.effort (default max),
so a high-frequency job like the watcher pays top-tier effort on every quiet
tick, which is the dominant per-turn cost on those jobs.

This adds agent.cron_effort (default medium) and applies it in
_build_options: a turn whose source is cron or hook uses cron_effort,
while interactive sources (web, telegram, wakeup) keep effort. The
selected value still flows through _effective_effort, so it is clamped to the
target model's effort ladder and stays safe on models with a shorter ladder.

The source-to-effort choice is factored into a small
_base_effort_for_source staticmethod so the policy is unit-testable in
isolation, mirroring the existing _effective_effort helper.

No migration needed: cron_effort defaults to medium, and a deployment that
wants the previous behavior can set cron_effort: max.

Test plan

• New unit tests in tests/test_engine.py:
  • test_base_effort_for_source: cron/hook map to cron_effort;
    web/telegram/wakeup/api map to effort.
  • test_base_effort_for_source_then_capped: routing composed with the
    model-cap clamp (Sonnet ladder, and pass-through for an uncapped model).
  • test_agent_config_cron_effort_default_and_override: from_dict default
    is medium, override respected, effort stays independent.
• python3 -m pytest tests/ -> 1355 passed, 2 skipped. The only 2 failures
  are pre-existing and unrelated (docker-mode detection tests that assume the
  host is not a container); they fail identically on main.