[Perf] Support tensor transmission between work and engine with zero-copy to improve efficiency

[sunlei1024]

Motivation

This pull request introduces tensor transmission support between the work and engine modules.
A zero-copy mechanism is implemented to eliminate unnecessary memory copying during data exchange, significantly improving transmission efficiency and reducing memory overhead.

---

Modifications

• Added support for transmitting tensor data between work and engine.
• Implemented zero-copy in the ZMQ-based communication layer to avoid redundant data copying.
• Optimized serialization/deserialization logic for high-performance data transfer.
• Ensured backward compatibility with the existing communication protocol and interfaces.

---

Accuracy Tests

This change only affects the communication layer and does not modify model computation logic.
Data consistency has been verified across multiple tensor types and shapes (bitwise equality confirmed).

---

Checklist

• Added appropriate PR tags ([Feature][Engine]).
• Code formatted and passed pre-commit checks.
• Verified zero-copy path correctness with tensors of different shapes and data types.
• No accuracy regression observed.
• Unit tests for tensor transmission to be added (pending review).

[paddle-bot]

Thanks for your contribution!