fix: cap evolve LLM summary tokens

[dfitz1138]

Respect EVOLVE_LLM_MAX_TOKENS when deriving evolve config from SkillClaw and cap session summarization requests at an OpenAI-compatible ceiling.

Tests: uv run pytest tests/test_evolve_config.py tests/test_skill_backend_config.py -q

[Upper9527]

The config bridge should respect EVOLVE_LLM_MAX_TOKENS, but the hard min(..., 8192) in the summarizer defeats that setting: an explicit value such as 32768 is still silently reduced to 8192. It also changes the established default from 100000 to 8192 only on the from_skillclaw_config path.

Please wire the environment value through consistently, then have the summarizer use the configured client limit (without a new hard-coded ceiling). Add coverage for an explicit value above 8192 and for the default matching the normal from_env path.