[fix] Device and format and implementation optimization

[KexinFeng]

1. Add the space char to the output
2. Add the test script of "huggyllama/llama-7b" for rolling_batch scheduler
3. Device setting is fixed. The following way will involve model copying, and cause OOM on huggyllama/llama-7b on 5.12xlarge.

        if self.device:
            self.model.to(self.device)

4. Efficiency optimization by avoiding list.pop(idx)