fix: set current CUDA device in _inplace_pin_memory function

[SongXiaoXi]

This pull request introduces a minor import adjustment and a fix to device handling when pinning memory for CUDA tensors. The main changes ensure that memory pinning happens on the correct CUDA device and resolve an import path issue.

CUDA device handling improvements:

• In checkpoint_engine/pin_memory.py, the current CUDA device is now explicitly retrieved and set (device_index = torch.cuda.current_device() and torch.cuda.set_device(device_index)) before pinning memory, ensuring that pinning occurs on the correct device. [1] [2]

Import path correction:

• In examples/update.py, the import for request_inference_to_update is fixed by importing it directly from checkpoint_engine instead of checkpoint_engine.ps, resolving a potential import error.