[onert] Introduce unified/single memory manager

[YongseopKim]

Current onert has a memory manager per an backend. This task introduces a single unified memory manager for all backends. This enables onert to make its own interface easy

Problems

• On arm arch, it doesn't seem to be not easy because of opencl memory concept. It should be invested more.
  • Therefore, it could be two memory pool for cpu and gpu.
• If each backend should own memory allocator?
  • While ruy has own memory allocator, xnnpack can take others' memory allocator.

In the future,

• Could use the better memory allocator like tcmalloc
• Could make onert to support memory bound feature.

References

• Spec 's 3.3Memory Model
• Relevant internal issue

When https://github.com/Samsung/ONE/issues/5006 is done, this task would start.

[periannath]

FYI, related article and paper.

• Optimizing TensorFlow Lite Runtime Memory : https://blog.tensorflow.org/2020/10/optimizing-tensorflow-lite-runtime.html
• Efficient Memory Management for Deep Neural Net Inference : https://arxiv.org/abs/2001.03288