On-device Implementation of ScaNN

ScaNN (Scalable Nearest Neighbors) is a method for efficient vector similarity search at scale. This is a simplified version of ScaNN that requires less resources to run and only for inference. There's no support for K-Means partitioning training and quantization training. It supports retrieval with the following features:

1. K-Means tree space partitioning.
2. Asymmetric Hashing (AH) quantization.
3. dot_product and squared_l2 distance measures. Note that for dot_product distance, we return the negative dot product. This is to ensure consistency with squared_l2 that smaller means closer.
4. Indexing new embeddings, including assigning them to closest partitions and AH quantize them.