Created two new modules for pycleora:

pycleora/search.py - ANNIndex class for approximate nearest neighbor search:
- ANNIndex(graph, embeddings, method="hnsw"|"brute") builds an ANN index
- .query(entity_id, top_k=10, exclude_self=True) returns results in same format
  as find_most_similar, with matching exclude_self parameter for API parity
- .query_vector(vector, top_k=10) for raw vector queries
- HNSW backend via optional hnswlib dependency (try/except ImportError pattern)
- Pure-numpy ball tree fallback when hnswlib is not installed
- Brute force method for exact baseline comparison
- Input validation for top_k parameter

pycleora/compress.py - Embedding compression utilities:
- pca_compress(embeddings, target_dim) - PCA via np.linalg.svd
- random_projection(embeddings, target_dim, seed=None) - fast alternative to PCA
- product_quantize(embeddings, num_subspaces, num_centroids) - standard PQ with
  k-means on subspaces, returns PQIndex with .reconstruct() and .search() methods
- Input validation for all parameters (target_dim, num_subspaces, num_centroids,
  max_iter, embeddings shape)

Updated pycleora/__init__.py to import both modules so users can access
pycleora.search.* and pycleora.compress.*.

Existing find_most_similar function is untouched for backward compatibility.
Result format (entity_id, index, similarity keys) matches find_most_similar exactly.

Replit-Task-Id: a2415004-3c28-4c96-b8e1-ed1365533871