Fold AgentModelFreeApproach and AgentSessionMixin into AgentBaseApproach

[yichao-liang]

Summary

Stacked on #192 (base empiric-prompt-submit-refs); review and merge after it.

AgentModelFreeApproach had become the base of every agent arm: AgentSimLearningApproach (and through it EMPIRIC and its model-based baselines) and AgentContinualModelFreeApproach, the model-free arm.
Its task solver only raised, and its name described neither role.
AgentSessionMixin had that class as its only host.

The two become one abstract base, AgentBaseApproach in approaches/agent_base_approach.py:

• the session machinery (tool-surface checks, session construction, queries) and the approach glue (option model, tool-context sync, checkpoints) live in one class;
• it defines no get_name, so it cannot be created by name, and the unused agent_model_free approach name is gone.

Behavior is unchanged: every arm inherits the same methods, _get_log_dir still creates the directory, and checkpoints keep their format.
Docs and comments that named the old classes are updated, including a few that still described the deleted Docker manager, agent explorers and phased learn hook.

Test plan

• yapf, isort 5.10.1, docformatter, mypy and pylint on this commit (sbatch replay of CI's static jobs).
• The 8 CI shards in the CI container: all pass except test_gather_timeout_returns_partial in shard 7, a timing flake independent of this change (reproduced and fixed in Wait for the fast rollout before timing gather's partial return #193).

🤖 Generated with Claude Code