Fix DeepSeek document block encoding

[zhexuany]

Summary

• Skip OpenAI file/input_file content blocks when the target provider or model resolves to DeepSeek.
• Preserve existing text extraction context for uploaded docx/documents instead of sending unsupported base64 file blocks to DeepSeek/LiteLLM.
• Add a regression test for deepseek-v4-pro routed through the OpenAI provider.

Test Plan

• cd packages/api && npx jest src/files/encode/document.spec.ts --runInBand
• cd packages/api && npm run build