Fix integration slow tests

[sgugger]

What does this PR do?

This PR fixes the following slow tests which are failing because of the change of beahvior in the Embeddings layer in PyTorh 1.8. This is done by adding an attention mask to ignore the padding token and checking a slice that does not contain the padding hidden states.

tests/test_modeling_albert.py::AlbertModelIntegrationTest::test_inference_no_head_absolute_embedding
tests/test_modeling_bert.py::BertModelIntegrationTest::test_inference_no_head_absolute_embedding
tests/test_modeling_bert.py::BertModelIntegrationTest::test_inference_no_head_relative_embedding_key
tests/test_modeling_bert.py::BertModelIntegrationTest::test_inference_no_head_relative_embedding_key_query
tests/test_modeling_convbert.py::ConvBertModelIntegrationTest::test_inference_masked_lm
tests/test_modeling_deberta.py::DebertaModelIntegrationTest::test_inference_no_head
tests/test_modeling_deberta_v2.py::DebertaV2ModelIntegrationTest::test_inference_no_head
tests/test_modeling_distilbert.py::DistilBertModelIntergrationTest::test_inference_no_head_absolute_embedding
tests/test_modeling_electra.py::ElectraModelIntegrationTest::test_inference_no_head_absolute_embedding
tests/test_modeling_squeezebert.py::SqueezeBertModelIntegrationTest::test_inference_classification_head

It also fixes

tests/test_modeling_mbart.py::MBartEnroIntegrationTest::test_enro_generate_batch

that was failing since the change in prepare_seq2seq_batch. For some reason a word is different but it was consistent in PyTorch 1.7/PyTorch 1.8 so I changed the desired target. @patil-suraj if you want to take a closer look, I'll leave it to you.

[LysandreJik]

This looks good to me, thanks for taking care of it!