fix: stop designed / multi-column resumes silently dropping sections in PDF extraction

[Sofa3xpert]

Fixes #400.

Problem

Designed resume templates (Canva, Adobe Express, ...) paint sidebar content over decorative vector panels. extract_text_from_pdf called to_markdown(doc, pages=pages), which treats those panels as significant images and drops every text line over them. Whole sections (skills, education, contact) vanish before the LLM sees them, silently deflating scores.

Before / after

Same persona as resume/sample.pdf (Barack Obama), rebuilt as a Canva two-column layout:

• default (main) -- 624 chars -- 6 of 14 fields recovered -- entire Contact / Skills / Education sidebar dropped (phone, email, all 4 skills, Harvard Law J.D.)
• this PR -- 900 chars -- 14 of 14 fields recovered

No regression on layouts that already worked: resume/sample.pdf extracts byte-identically (906 to 906 chars). Verified against main @ 70fd3ea.

Fix

One line in pdf.py, pass ignore_graphics=True to to_markdown:

resume_text = to_markdown(
    doc,
    pages=pages,
    ignore_graphics=True,
)

Vector graphics in a resume are decorative; they must not suppress text. (ignore_graphics also skips markdown table formatting, but table text is still extracted, which is what matters for parsing.)

Test

Adds tests/pdf_extraction_test.py, a hermetic PDF to Markdown smoke test (no API key or network): it builds the Canva-style fixture and asserts the sidebar survives extraction. Fails on main, passes with this change. Black-clean; pytest-discovered and runnable standalone.

Run python tests/pdf_extraction_test.py (or pytest tests/).

Note: named *_test.py rather than test_*.py because the repo's .gitignore ignores test_*.py.