Being Comes from Not-being: Open-vocabulary Text-to-Motion Generation with Wordless Training