TUI support for mtmd API

[aittalam]

• implements multimodal (image) support in the TUI chatbot using the new mtmd API, replacing the previous stub that returned "image processing not yet implemented"
• aligns with llama.cpp server's multimodal handling for consistency

Closes #851

Changes

• added eval_mtmd_chunks() helper that evaluates tokenized chunks using mtmd_helper_eval_chunk_single() (matching the approach used by llama.cpp server)
• refactored eval_string() to:
  • scan input for data:image/... URIs
  • collect images as mtmd::bitmap objects
  • replace data URIs with <__media__> markers
  • tokenize text + images together via mtmd_tokenize()
  • evaluate the resulting chunks

Testing this PR

• Build with make -j8
• Run with a multimodal model: ./llamafile -m <model.gguf> --mmproj <mmproj.gguf>
• Use /upload <image_path> to upload an image and verify it processes correctly (note: spaces are not allowed in <image_path> yet)
• Test text-only prompts still work (fast path)
• Test multiple images in a single prompt