Port FractRAG multi-scale retrieval as RetrieverPlugin

[deesatzed]

Summary

• Port fractLrag's three-level self-similar index (sentence / paragraph / document) with derivative signals and retrieve_adaptive into CAM-RAG as cam_rag.retrieval.fractal.
• Add opt-in FractalRetrieverPlugin (name="fractal") that implements RetrieverPlugin.index / retrieve. It is not enabled in the default RAGAppSpec; callers pass it via retriever_plugins.
• HashEmbedding works without sentence-transformers or GPU. CAM-RAG HashEmbeddingBackend is also accepted.

Test plan

• HashEmbedding fractal retrieve returns ranked RetrievalResults
• Plugin satisfies RetrieverPlugin protocol (isinstance, name, empty query, pipeline RRF when requested)
• Existing CAM-RAG retrieval tests still pass
• Skip live PubMed / BGE / sentence-transformers paths

python -m pytest tests/test_fractal_plugin.py tests/test_fractal_chunking.py tests/test_fractal_rrf.py tests/test_fractal_query_classifier.py tests/test_fractal_virtual_paragraphs.py tests/test_fractal_weighted_pooling.py tests/test_fractal_diversity_scoring.py tests/test_fractal_benchmark.py tests/test_fractal_storage.py tests/test_fractal_metadata.py tests/test_retriever_plugin.py tests/test_retrieval_dense.py tests/test_retrieval_sparse.py tests/test_retrieval_fusion.py -q
# 173 passed

Not ported

• legacy/ (doctronic / xAI demos) — excluded by request
• medical_corpus.json / medical_queries.json — corpus dumps; corpus-only classifier/chunking cases omitted
• Original test_benchmark_fractal_vs_flat.py imported legacy engines; rewritten against unified FractalRAG + HashEmbedding
• Live BGE / PubMed / SentenceTransformerEmbedding load paths — optional, not required for HashEmbedding tests