fix(ai): expose Kimi Coding K3 effort levels

[dannote]

Summary

• expose low, high, and max thinking levels for Kimi Coding K3
• keep direct Moonshot K3 on its separate max-only contract
• update thinking-level regression coverage for both endpoints

Kimi Code now documents low, high, and max for k3: https://www.kimi.com/code/docs/en/kimi-code/models.html#model-overview

This is a follow-up to #6737. The max-only metadata added there matched Kimi Code's documented capability at the time, but the subscription endpoint has since added low and high.

Tests

• node node_modules/vitest/dist/cli.js --run test/supports-xhigh.test.ts from packages/ai
• npm run check
• ./test.sh

[badlogic]

cheers, fixed here 8b93737