Perf boost for batch embeddings

generate_embeddings_batch now checks MemoryStore per text before encoding, deduplicates repeated strings, and encodes only unique misses in chunks of 32. Cache hits reuse the stored vector; fresh computes are written back. Input order is preserved. This helps in consolidation flows where the same key string recurs across the batch.

• Cache lookup in batch path with de-dup

• Checks MemoryStore per text before encoding

• Deduplicates repeated strings

• Encodes only unique misses in chunks of 32