fix: run real ingestion during token estimation

[sumukshashidhar]

Summary

Fixes the token estimation logic to use actual document content instead of hardcoded estimates.

Changes

• Replace hardcoded 1K token estimate for PDFs with actual document extraction
• Use MarkItDown (without LLM) to extract content from PDFs, docx, etc.
• Simulate chunking to get accurate chunk counts
• Token estimates now based on real tiktoken encoding of actual content

Before

Source Documents:
  Files: 1
  Estimated tokens: 1K  # hardcoded guess

After

Source Documents:
  Files: 1
  Estimated tokens: 33.8K  # actual extracted content

Single Hop Question Generation:
  Input: 38.8K, Output: 7.5K, Calls: 5  # based on real chunk count

Testing

• All 91 tests pass
• Verified with yourbench estimate example/default_example/config.yaml
• Verified with yourbench estimate config.yaml in harry_potter_quizz directory