Conformance test for weight compression

[ljaljushkin]

Changes

Added tests for accuracy and duration of weight compression on tiny-llama-1B

Reason for changes

catch bugs with accuracy (e.g. difference after changing rounding - #2420) and with duration of weight compression

Related tickets

130630

Tests

• local run of weight compression conformance tests
• PTQ conformance tests in CI (build 273)
• CI scripts should be adopted for a new location of reference.yaml
• html generation should accept new names in the table.

Total execution time for tests w/ or w/o FP32 option is ~20/~40 minutes.
Model download takes extra 20 minutes, but with caching models on share #2427
the download won't happen.

[codecov]

Codecov Report

Attention: 1 lines in your changes are missing coverage. Please review.

Comparison is base (c0e51f8) 91.02% compared to head (837d695) 84.86%.
Report is 8 commits behind head on develop.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2425      +/-   ##
===========================================
- Coverage    91.02%   84.86%   -6.17%     
===========================================
  Files          496      497       +1     
  Lines        45079    45222     +143     
===========================================
- Hits         41035    38379    -2656     
- Misses        4044     6843    +2799     

Files	Coverage Δ
...n/algorithms/weight_compression/mixed_precision.py	46.66% <0.00%> (-51.43%)	⬇️

... and 64 files with indirect coverage changes

Flag	Coverage Δ
COMMON	44.23% <0.00%> (-0.37%)	⬇️
ONNX	35.01% <0.00%> (-0.08%)	⬇️
OPENVINO	∅ <ø> (∅)
TENSORFLOW	30.19% <0.00%> (-0.07%)	⬇️
TORCH	66.16% <0.00%> (+2.48%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Components	Coverage Δ
common	92.71% <ø> (-0.66%)	⬇️
torch	93.38% <ø> (+0.20%)	⬆️
tensorflow	93.99% <ø> (ø)
onnx	98.01% <ø> (ø)
openvino	26.03% <ø> (-68.42%)	⬇️
ptq	68.40% <0.00%> (-20.68%)	⬇️

[ljaljushkin]

job 4

report looks good.
accuracy results for data-aware methods are not stable. will investigate and fix later