[WeightCompression] Statistics caching

[kshpv]

Changes

Add statistics saving and loading for the WeightCompression algorithm:

1. Statistics are cached for all configurations such as:
   awq = True, scale_estimation=True with all type of sensitivities.
2. Then the statistics are dumped in a directory which can be reused for any weights_compression() configuration.

The example for tinyllama was updated with this functionality.

More changes:

1. Make all statistics used in WeightCompression aligned with TesnorStatistics from nncf/experimental/common/tensor_statistics/statistics.py
2. Extend StatisticsAggregator by the logic of loading and saving statistics.
3. Dumping statistics is done using pickle and gzip. Serialization methods were added for Tensor.
4. Introduced statistics_serializer and statistics_validator to handle the statistics loading/dumping to the file.

Statistics sizes

Model	subset size	statistics directory size	statistics collection time
tinyllama	128	100 MB	61 sec
Phi-3-mini-4k-instruct	128	258 MB	51 sec
ruDialoGPT-medium	128	80 MB	9 sec
llama-3.1-8b	128	393 MB	95 sec

Reason for changes

Speed up compression configuration finding.

Related tickets

153129

Tests

Test coverage were extended by tests on statistics_serializer, statistics_validator, StatisticsAggregator and on WeightCompression algorithm with the proposed functional.

[ljaljushkin]

By taking into account that gptq could be true and lora_correction would be great to test as well, do you see any opportunities to further reduce the time? Maybe smaller ratio, e.g. [0.2, 0.4]?
What is the most consuming part?

[daniil-lyakhov]

TensorCollector and co LGTM

[alexsu52]

I marked this PR as request changes because the implementation is not optimal.

[alexsu52]

LGTM, please address my minor comments