Do one-time shuffle of document codes so that it reflect insertion order, making it Xet friendly, and add sources up to Sept 30

Add empty.faiss and reduced ids.parquest

Update trained index because Stella doesn't use truncated vectors, cut cluster count to save memory (normalized, OPQ96_384,IVF65536,PQ96)

Add trained index to hopefully its final place (truncated to length-384, normalized, OPQ96,IVF131072,PQ96)