Smarter memory allocation and reuse for DLK runtime

[primenumber]

• Reduce memory consumption
• Erase useless memory copy
  In dlk runtime, buffer size are known in compile time, we can optimize memory usage and memory transfer.

[primenumber]

Some operations require temporary buffer to compute.
In current implementation, temporary buffer are managed by global std::unique_ptr<T[]> to avoid memory allocation in inference time.
It prevent concurrent inference and increase memory consumption.
We should consider to manage these buffer in Network.

[primenumber]

Temporary buffer is now allocated in Network::init except RunTCA.