feat: refactor PDF extraction to caption-based screenshots, add upvote refresh, clean up UI

- PDF extractor: rewrite from embedded bitmap extraction to caption-based
  page region screenshots. Finds Figure/Table captions via regex,截取上方/下方
  page region, handles compound figures and vector graphics.
- Upvote refresh: new crawler.refresh_upvotes() re-fetches upvotes for recent
  N days without inserting new papers. Scheduler runs daily 30min after pipeline.
- Admin: add /admin/refresh-upvotes endpoint and dashboard button.
- UI: remove date quick nav, show upvote update time on detail/card pages,
  clean up CSS date-chip styles.
- Utils: add recent_date_strs() helper.

app/templates/partials/paper_card.html

@@ -20,7 +20,7 @@
         {% endif %}
       </a>
     </h2>
-    <span class="paper-upvotes">👍 {{ paper.upvotes }}</span>
+    <span class="paper-upvotes" title="数据更新于 {{ paper.crawled_at.strftime('%m-%d %H:%M') if paper.crawled_at else '' }}">👍 {{ paper.upvotes }}</span>
     {% if variant == 'search' and distances and paper.arxiv_id in distances %}
     <span class="similarity-score" title="语义相似度距离">
       🎯 {{ "%.3f"|format(distances[paper.arxiv_id]) }}