Unshackling Context Length: An Efficient Selective Attention Approach through Query-Key Compression