Fix error when cls_pooling="mean" or cls_pooling="max" for TransformerDocumentEmbeddings

[fkdosilovic]

• Refactor "cls" pooling option into a function.
• Return result from document_mean_pooling and document_max_pooling functions.
• Add unit tests for testing each cls pooling option.
• Refactor test_default_embeddings_stay_the_same_after_saving_and_loading and test_embeddings_stay_the_same_after_saving_and_loading in BaseEmbeddingsTest to use torch.allclose

Fixes #3552.

[fkdosilovic]

I thought that the two failing tests (test_default_embeddings_stay_the_same_after_saving_and_loading for TransformerWordEmbeddings and TransformerDocumentEmbeddings) were somehow due to my changes, but this doesn't seem to be the case.

When running tests locally, every test passes. Then I ran the ci.yml using act to simulate the CI job locally. I ran the act -j 'test' and got those tests to fail on master.

Looking at the output, the problem was in comparing the floating point values (tensors) with == instead of using something as torch.allclose.

I changed this == to use torch.allclose in both test_default_embeddings_stay_the_same_after_saving_and_loading and test_embeddings_stay_the_same_after_saving_and_loading.

Did you encounter similar problems before?

[helpmefindaname]

Thank you for the bugfix @fkdosilovic
I will merge as soon as the tests complete