Fix safetensors index to match actual 3-shard + lm_head layout

Add lm_head.safetensors for logit scoring + update model card with verified results

Add model card with tags

Add Qwen3-Reranker-8B converted to MLX fp16 via mlx-embeddings

initial commit