Add generic KG embedding module with RRF hybrid search support

New file: embed_kg.py (442 lines)

Purpose: Generic embedding-based knowledge graph search supporting three modes:
1. Embedding-only: Dense vector similarity using KaLM
2. BM25-only: Keyword matching (via ARK's GraphIndex)
3. Hybrid: RRF fusion combining both for best of both worlds

Key Components:
- clean_query(): Copied from ark/src/core/index.py to ensure embedding/BM25 consistency
  - Cleans node names identically across both search methods
  - Comment explains why copy was necessary (SLURM -m flag module import issues)

- KGEmbedder class:
  - load_nodes(): Loads KG parquet + cleans names to match BM25 index
  - embed_all_nodes(): One-time operation using KaLM encode_document()
  - embed_query(): Per-request encoding using KaLM encode_query()
  - search_embedding_only(): Dense similarity search returning top-k results
  - reciprocal_rank_fusion(): Combines embedding + BM25 via RRF formula
    - score = 1/(rrf_k + rank) for each method, then summed
    - Both rankers weighted equally (symmetric fusion)
    - Returns fused_score + metadata (embedding_rank, bm25_rank, individual scores)
  - save_embeddings() / load_embeddings(): Disk persistence for efficiency

- CLI interface:
  - --graph-path: KG directory with nodes.parquet
  - --output-path: Embeddings save location (default: graph_path/embeddings_kalm.npy)
  - --batch-size: Embedding batch size (default: 256)
  - Validates inputs, runs 3-step pipeline (load → embed → save)

Compatible with: PRIME, MAG, AMAZON, OptmusKG, any ARK-compatible KG

Impact: Enables exploring hybrid search strategies (RRF fusion) without modifying ARK core, prepares for search-mode parameter integration in Phase 1A.

Author: Reza Shamji