perf: batch=1 优化减少延迟

- settings: batch_size=41
- tensorrt_engine: BATCH_SIZE=41
- preprocessor: 移除 padding 逻辑，直接 batch=1
- 预处理延迟从 17ms  5ms

core/tensorrt_engine.py

@@ -40,29 +40,22 @@ class HostDeviceMem:
 
 def pad_to_batch4(frames: List[np.ndarray]) -> np.ndarray:
     """
-    Padding 到 batch=4，重复最后一帧
+    Padding 到 batch=N，重复最后一帧（已弃用，改用 batch=1）
     
     Args:
         frames: list of [3, 480, 480] numpy arrays
     
     Returns:
-        np.ndarray: [4, 3, 480, 480]
+        np.ndarray: [N, 3, 480, 480]
     """
     if len(frames) == 0:
         raise ValueError("Empty frames list")
     
-    if len(frames) == 4:
-        return np.stack(frames)
-    
-    pad_frame = frames[-1].copy()
-    while len(frames) < 4:
-        frames.append(pad_frame)
-    
     return np.stack(frames)
 
 
 class TensorRTEngine:
-    """固定 batch TensorRT 引擎 (batch=4, FP16, 3×480×480)
+    """TensorRT 引擎 (batch=1, FP16, 3×480×480)
     
     特性：
     - Buffer Pool: bindings 只在 init 阶段分配一次
@@ -70,7 +63,7 @@ class TensorRTEngine:
     - Async API: CUDA stream + async memcpy + execute_async_v2
     """
     
-    BATCH_SIZE = 4
+    BATCH_SIZE = 1
     INPUT_SHAPE = (3, 480, 480)
     
     def __init__(self, config: Optional[InferenceConfig] = None):