📝 docs: enhance RAG pipeline documentation with structured extraction details

## Implementation Phases

### Phase 0: Compatibility and Unification Design

- Defined reusable structured models (`ExtractedTableElement`, `ExtractedFigureElement`, `ExtractedEquationElement`)

- Added extraction feature toggles to `LightRAGConfig`

- Established provider compatibility matrix for vision capabilities

### Phase 1: Shared Structured Extraction Core

- Created unified extraction function returning `StructuredExtractionResult`

- Integrated with `processor.py`, `convert.py`, and `summarize.py`

- Established consistent `.ai/` artifact layout

### Phase 2: Table Preservation

- Converted `result.tables` into structured elements with IDs, page numbers, dimensions

- Added stable table markers in markdown output

- Created JSON sidecars for machine-readable structure

### Phase 3: Figure/Image Extraction

- Persisted extracted figures under `.ai/figures/`

- Implemented caption matching heuristics

- Added cached figure description via `LiteLLMClient` with graceful non-vision fallback

### Phase 4: Equation Handling and Structural Chunking

- Detected and preserved equation blocks (`$$`, `\[ ... \]`, `\begin{equation} ...`)

- Introduced structural chunking respecting table/figure/equation boundaries

- Standardized metadata propagation through `TDocRAG.insert()`

### Phase 5: Single-Command Query Enhancement

- Maintained `tdoc-crawler ai rag query` as the single query command

- Improved retrieval context using enriched chunks from all element types

- Enhanced citation formatting with element type and location

## Validation

All PRs validated through:

- Unit tests: `uv run pytest packages/3gpp-ai/tests/test_*.py -v`

- End-to-end workflow: workspace create → add-members → process → rag query

- Artifact verification: `.ai/` folder contains markdown, JSON sidecars, and figures