Fix paper review JSON parse when Hub reasoning models empty content

[znsoftm]

Hub auto reasoning models can return HTTP 200 while spending max_tokens on reasoning_content and leaving message.content empty. extractChatContent previously returned that empty string, so parseReviewAnalysis failed with unexpected end of JSON input on POST /papers/review/{id}/generate.

Changes

• Raise reviewMaxTokens from 2500 to 12288 so reasoning plus a full four-section Chinese JSON review can fit.
• Harden extractChatContent:
  • Clear error when content is empty (empty chat content (finish_reason=…))
  • Recover an embedded JSON object from reasoning_content / reasoning when present
• parseReviewAnalysis rejects empty input with empty analysis JSON instead of a raw JSON EOF.
• Tests for empty content, reasoning-only Hub 200 shapes, JSON recovery, and the new token budget.

Base

Stacked on cursor/papers-translate-concurrency-7d3b (merge target of the 解读 review feature, PR #11). main does not yet include papers review generation.

Tests

go test ./... passes, including:

• TestExtractChatContentShapes
• TestGenerateReviewOpenAIReasoningOnly
• TestGenerateReviewOpenAIEmptyContentError
• TestReviewMaxTokensBudget