feat: support DeepSeek Vision for PDF resume import

[gziyin]

Summary

• Add DeepSeek Vision support to PDF resume import using the official experimental deepseek-v4-flash-vision-exp model.
• Route PDF imports through DeepSeek when it is the selected AI provider, while preserving Gemini as the backward-compatible default.
• Share proxy initialization between Gemini and DeepSeek and map network/upstream failures to safe client responses.
• Add 20 focused tests for provider normalization, payload construction, JSON parsing, proxy initialization, and DeepSeek error handling.

Problem

PDF resume import was hard-wired to Gemini. Users who selected and configured DeepSeek for AI features still had to configure a Gemini key before importing a PDF, so the import behavior did not match their provider selection.

Implementation

• Extend /api/resume-import with an optional modelType discriminator. Missing values still select Gemini for existing callers; unsupported values return HTTP 400.
• Send PDF page images as OpenAI-compatible image_url data URLs to DeepSeek's Chat Completions endpoint using deepseek-v4-flash-vision-exp and JSON output mode.
• Reuse one resume schema prompt and one tolerant JSON-object parser across both providers.
• Keep the successful response contract unchanged as { resume }.
• Move the existing Gemini proxy setup into a shared server module and initialize it before either provider makes an outbound request.
• Return fixed public messages for DeepSeek failures while retaining upstream details only on the server-side error object.
• Update Chinese and English import guidance to refer to the selected AI provider.

DeepSeek Vision request shape and limits follow the official documentation:
https://api-docs.deepseek.com/guides/vision/

Compatibility

• Requests without modelType continue to use Gemini.
• Existing PDF rendering and resume persistence flows are unchanged.
• Selecting DeepSeek uses the DeepSeek API key; other providers retain the existing Gemini PDF-import fallback.
• API keys, proxy URLs, and raw upstream response bodies are not returned to the client.

Verification

• pnpm test:resume-import: 20/20 passed
• pnpm build: client and SSR builds passed
• pnpm exec tsc --noEmit: the repository still reports 41 pre-existing diagnostics; changed files report 0 diagnostics
• Chinese and English locale JSON parsed successfully
• git diff --check passed
• Changed-file secret and local-path scan returned no matches

The existing DEFAULT_TEMPLATES circular chunk warnings and large bundle warnings are unchanged.

Manual validation still required

• Import a real multi-page PDF with a valid DeepSeek API key and verify the generated resume fields.
• Repeat the same browser flow with Gemini to confirm backward compatibility.

This PR is opened as a draft until those credential-dependent checks are completed.