Don't Look Twice: Faster Video Transformers with Run-Length Tokenization