fix: increase test connection max_tokens to satisfy GPT-5 minimum

[devmarkpro]

Summary

• GPT-5 series models (e.g., GPT-5.4) require max_output_tokens >= 16, but the "Test Connection" button in AI settings was sending max_tokens: 10
• This caused a 400 error: Invalid 'max_output_tokens': integer below minimum value. Expected a value >= 16, but got 10 instead.
• Increased to 20, which is compatible with both GPT-4 (min 1) and GPT-5 (min 16) model families
• The production usage in NotificationService (which uses maxTokens: 300) is unaffected

Test plan

• Configure a GPT-5 series model (e.g., gpt-5.4) in Settings → AI and click "Test Connection" — should succeed
• Configure a GPT-4 series model and click "Test Connection" — should still succeed (no regression)

[renjfk]

Thanks!