How attention sinks keep language models stable