Add matplotlib implementation as tex rendering fallback, and subsequent improvements to text visualization and output sizing

[zzyking]

Summary (diff vs previous)

• Added a first-class matplotlib/CV fallback for text rendering (escapes $, consistent spacing/padding) so environments without texlive-xetex/sudo still produce readable outputs; docs now mention the fallback and Pillow was added for DPI-aware saves.
• Set default render DPI to 200 and switched to PIL-based saving to speed up renders while keeping clarity.
• Tightened default viz sizing (width 800) with aspect-ratio helpers; video outputs scale per-frame to avoid too wide composite image-text outputs, which observably slows down processing.
• Removed fitz from requirements because pymupdf can fully replace it.

Motivation

• texlive-xetex is bulky and often unavailable on clusters without sudo. The fallback keeps text viz usable and quicker, while DPI/sizing tweaks prevent oversized outputs.

Testing

• Selected one visualization for each settings in demo.py (qwen2vl, internvl3, qwen2vl_video from top to bottom)