PoC of lowering compilation time using Python threading

[p-wysocki]

Changes

• implemented Python threading in order to speed up model compilation
• removed the need to serialize the model after compilation
• using https://github.com/openvinotoolkit/openvino_notebooks/blob/main/notebooks/122-quantizing-model-with-accuracy-control/122-speech-recognition-quantization-wav2vec2.ipynb as a benchmark, the step:
  INFO:nncf:Calculating ranking score for groups of quantizers has been shortened from ~4 minutes to a little over 2 minutes
• of the remaining 2 minutes, the compilation takes only ~12 seconds on my machine

Related tickets

119274

[alexsu52]

A general comment is that the proposed solution is not memory optimal. What I mean is that this solution requires storing as many copies of the model in memory as there are groups in groups_to_rank. For some models, the number of groups for ranking is in the hundreds. Probably, algorithm can be crashed by memory for huge models.