[codex] Fix AI SPI classloader fallback

[kobewl]

Summary

• load AI SPI implementations from the thread context class loader first, then fall back to the SPI interface class loader
• apply the same fallback behavior to both AIServiceFactory and AIConfigRegistry`n- add regression tests that simulate a context class loader which cannot see the SPI resources

Root cause

ServiceLoader was only using the current thread context class loader. In thread-pool or Spring-style execution environments, that loader may not be able to resolve META-INF/services, which leaves the AI provider/config registries empty and causes Unsupported model errors.

Validation

• mvn -pl hutool-ai -am -Dsurefire.failIfNoSpecifiedTests=false -Dtest=AISpiClassLoaderCompatibilityTest,AIServiceFactoryTest test (run with JDK 8)

[looly]

这块乱码了，我手动修复了。