Preserve content list aliases in multimodal chunks

[FU-max-boop]

Summary

This PR makes direct content_list multimodal processing preserve common alias fields instead of silently dropping or misformatting them.

Changes:

• preserve each multimodal item's original content_list index after text items are separated, so context lookup still points at the original surrounding items
• normalize string/list caption and footnote fields before joining them into image/table chunks
• read table content from table_body, table_data, or text
• read equation content from text, latex, or equation, and infer LaTeX format when only latex is provided
• add an offline regression test for these alias-handling cases

Why

The direct content_list path can receive pre-parsed multimodal items whose field names differ slightly by source. In that path, a table that only has table_data can produce an empty table body, an equation that only has latex can lose its formula text, and a string caption can be joined character-by-character.

The same path also separates text and multimodal items before later context extraction. Without preserving the original item index, context lookup can point to the compacted multimodal list position instead of the original content_list position.

Validation

From the repo root:

python3 -m py_compile raganything/utils.py raganything/processor.py raganything/modalprocessors.py tests/test_content_list_alias_handling.py
python3 -m unittest tests/test_content_list_alias_handling.py
git diff --check

The targeted unittest covers:

• original content_list index preservation
• string caption normalization for table chunks
• table_data alias preservation
• latex equation alias preservation

Scope / limits

This PR does not change parsers, model calls, retrieval behavior, or prompt templates. It only normalizes already-parsed multimodal item fields before building prompts/chunks.