Implement OpenAI-compatible generation for remaining AI stubs

[ifanatics-media]

Several AI adapter packages still return stub placeholder text even though the providers expose OpenAI-compatible chat-completions APIs:

• packages/ai/kimi
• packages/ai/cohere
• packages/ai/novita
• packages/ai/venice
• packages/ai/parasail

Current behavior means generate() succeeds with [stub ... integration not yet implemented] instead of calling the provider, so users can configure API keys and still get fake output.

Expected behavior:

• use each provider's documented OpenAI-compatible chat-completions endpoint
• preserve dry-run behavior
• forward system/user messages, maxTokens, temperature, and opts.extra
• map choices[0].message.content, response model, and token usage
• cover request shape and error handling in tests