[Backend] TRT backend & PP-Infer backend support pinned memory (#403)

* TRT backend use pinned memory

* refine fd tensor pinned memory logic

* TRT enable pinned memory configurable

* paddle inference support pinned memory

* pinned memory pybindings

Co-authored-by: Jason <jiangjiajun@baidu.com>

fastdeploy/backends/tensorrt/utils.h

@@ -206,6 +206,8 @@ class FDGenericBuffer {
 };
 
 using FDDeviceBuffer = FDGenericBuffer<FDDeviceAllocator, FDDeviceFree>;
+using FDDeviceHostBuffer = FDGenericBuffer<FDDeviceHostAllocator,
+                                           FDDeviceHostFree>;
 
 class FDTrtLogger : public nvinfer1::ILogger {
  public: