Add agent.model_effort: per-model replacement for agent.effort

[serxa]

What

agent.effort applies to every model. There was no way to run one model at a different effort than the rest, short of changing the default for all of them.

Change

New agent.model_effort: a map from model-name substring to effort level that replaces agent.effort for matching sessions.

agent:
  effort: max
  model_effort:
    opus-5-5: high

• Keys are case-insensitive substrings of the session's model name, checked in order; the first match wins. Models that match no key keep agent.effort.
• It only replaces the agent.effort default. Cron and hook turns keep agent.cron_effort, and an explicit per-job (jobs.yaml) or per-run effort still wins.
• An entry with an unknown level is dropped with a warning, so a typo falls back to agent.effort instead of sending no effort at all.
• It is read through the same live config reference as agent.effort, so a reload applies it to clients built afterwards.
• docs/config.md documents it, together with agent.effort and agent.cron_effort, which the Agent table was missing.

Tests

• tests/test_engine.py: matching, case, order, validation and defaults.
• tests/test_model_effort.py: the effort on the SessionSpec built by _get_or_create_client for web, Telegram and cron turns, for a model with no entry, and with an explicit override.