把20步的ddim更换为10步的dpm-solver++，采样时间快40%，效果基本一致

[sah1234567]

唯一的问题是在lipsync_pipeline.py中对每个视频块采样前必须重置调度器状态计数，不然会报index超过size的错误。
可以在以下位置，加入以下内容：
`
for i in tqdm.tqdm(range(num_inferences), desc="Doing inference..."):
# ============ 添加这一段调度器状态重置代码 ============
# Reset scheduler state for each inference chunk to avoid index out of bounds
self.scheduler.set_timesteps(num_inference_steps, device=device)

        # Reset DPMSolverMultistepScheduler internal state
        if hasattr(self.scheduler, 'model_outputs'):
            self.scheduler.model_outputs = [None] * self.scheduler.config.solver_order

        # Use set_begin_index method to reset step tracking
        if hasattr(self.scheduler, 'set_begin_index'):
            self.scheduler.set_begin_index(0)

        # Ensure timesteps are properly set for this iteration
        timesteps = self.scheduler.timesteps
        # =====================================================

`