Use AvgPool-14 in PT FE

[p-wysocki]

Details:

• Extend PT FE with AvgPool-14
• Move the count_include_pad workaround subgraph from PT FE to a downgrade transformation
• The subgraph in downgrade transformations is defined in https://github.com/openvinotoolkit/openvino/blob/master/src/common/transformations/src/transformations/op_conversions/convert_avgpool_downgrade.cpp#L52
• The new RoundingType::CEIL_TORCH has impact only on the output shape
• Out of two tests which previously have been xfailed due to output shape mismatch one is still failing, but now due to accuracy validation failure
• No old tests were broken by upgrading to V14, so it seems the subgraph has issue with this edge case specifically
• The edge case is connected to count_include_pad:True, and pad value, depending on the latter FP16 test case sometimes passes, but FP32 always fails
• Since CPU Plugin is calling oneDNN AvgPool implementation there may be some difference in results calculation or an issue with parameter preprocessing before calling oneDNN function
• Continued in XXXXX

Tickets:

• CVS-133929

[mlukasze]

setting do not merge label temporary, as we found additional issue with operator and it's need to be discussed before we enable it for the frontend

[mvafin]

With merge of MaxPool-14 we had performance regression on CPU before it was implemented in plugin and on GPU. Can we test that this change doesn't degrade performance?

[github-actions[bot]]

This PR will be closed in a week because of 2 weeks of no activity.

[eaidova]

@p-wysocki please make sure that updasted IR works on all devices including NPU and downgrade transformation where op is not implemented applied. We still see issue for maxpool (during 2 releases for NPU), not lower amount computer vision models will be affected by switching by avgpool