Upsert pgvector records

Current implementation [deletes all of chunks before creating](https://github.com/aws-samples/bedrock-claude-chat/blob/main/backend/embedding/main.py#L49-L82) to keep idenpotency. But upsert operation is better in terms of performance especially for large dataset.
Primary key may need to be changed as: `s3_uri + ':' + chunk_num `

Start with backend/embedding/main.py, especially lines 49-82, to understand how chunks are currently deleted and recreated. Then inspect the existing pgvector record structure and primary-key handling. Done means replacing the full deletion flow with an upsert that remains idempotent for large datasets.

Written by the indexing model from the issue text.