feat: support max reasoning effort

[anantgar]

Summary

• Accept max through CLI flags, config, and TypeScript/Python types.
• Forward max to Codex, Claude, OpenCode, and Pi using each backend’s native setting.
• Encode Cursor max as [effort=max] while preserving existing legacy variant mappings.
• Keep usage pricing correct for parameterized Cursor model names.

Verification

• npm run build passes.
• Focused max/agent tests: 17 passed.
• Live headless probe: ChatGPT-authenticated Codex with gpt-5.6-sol and max returned OK.
• Full npm suite: 454 passed; 5 existing Modal tests fail on invalid pax header in the environment.
• Python suite: 66 passed; 1 packaging test requires the missing build module.

Effort support remains model/backend-specific. Headless forwards the selected value and lets the native backend report unsupported combinations rather than silently downgrading them.