Non-adversarial training of Neural SDEs with signature kernel scores