[QDP] gpu tensor zero-copy via dlpack protocol

[rich7420]

Summary

This issue asks gpu tensor support using dlpack protocol for zero-copy data transfer. This is critical for training loop performance where tensors are already on gpu.

• extend encode_tensor() to handle gpu tensors
• add dlpack extraction
• add gpu pointer validation
• add stream synchronization
• add direct gpu processing

If anything important is missing, please leave a comment.