test(agents): stabilize agent session timing assertions

[shizhigu]

Summary

• Split wall-clock event timestamp checks from deterministic duration and playback-position checks.
• Allow a small extra late margin for created_at event emissions under CI scheduler load while keeping early events strict.
• Preserve the wider early tolerance that test_generate_reply already used for its late conversation events.

Tests

• uv run ruff check tests/test_agent_session.py
• uv run ruff format --check tests/test_agent_session.py
• uv run pytest tests/test_agent_session.py::test_tool_call tests/test_agent_session.py::test_interruption_before_speaking tests/test_agent_session.py::test_generate_reply tests/test_agent_session.py::test_preemptive_generation -q
• uv run pytest tests/test_agent_session.py -q
• git diff --check

Claude reviewed the diff after the tolerance split and found no blocking issues.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.

[shizhigu]

Closing this separate test-only PR because the same stabilization commit is now folded into #5593, which was repeatedly hitting the same AgentSession timing assertion in CI. Keeping the docs fixes and the CI-unblocker together avoids making reviewers coordinate two dependent PRs.