Efficient PyTorch Tensor Sharing in Multi-process Environment

Add SharedTensorPipe, a wrapper optimized for torch.Tensor sharing.

Improve throughput by 30%+ (360msg/s -> 510msg/s on local test case) compared to naive mp.Pipe by tensor coalescing and shared-memory reusing.