Pie-modules

* encapsulate `get_pooler_and_output_size` in `SequenceClassificationModelWithPoolerBase` (124) ArneBinder
* add parameter `max_num_negatives` to `add_negative_coref_relations()` (123) ArneBinder
* `enforce_same_original_doc_id` parameter & improved text pair metadata (121) ArneBinder

:rocket: Features

* add parameter `enforce_different_original_doc_id` to `add_negative_coref_relations()` (122) ArneBinder

:boom: Breaking Changes

* `RETextClassificationWithIndicesTaskModule`: auto-align argument token spans (118) ArneBinder

:rocket: Features

* add parameter `random_seed` to `add_negative_coref_relations` (117) ArneBinder

* implement `TaskModuleWithDocumentConverter` (114) ArneBinder
* implement `SequenceClassificationModelWithPoolerBase` (112) ArneBinder

:rocket: Features

* SequenceClassificationModelWithPoolerBase: dedicated `setup_base_model` method (115) ArneBinder
* task: cross-text-coref (110) ArneBinder
* implement `allow_discontinuous_text` for `RETextClassificationWithIndicesTaskModule` (109) ArneBinder

* outsource `RelationStatisticsMixin` (from `RETextClassificationWithIndicesTaskModule`) (111) ArneBinder

:rocket: Features

* add `add_global_attention_mask_to_input` to `RETextClassificationWithIndicesTaskModule` (108) ArneBinder
* RE taskmodule: improve marker handling (107) ArneBinder

* add flair SegtokSentenceSplitter (105) RainbowRivey

:beetle: Fixes

* do not create negative relations if `create_candidate_relations=True` (`RESpanPairClassificationTaskModule`) (103) ArneBinder
* fix decoding of `RESpanPairClassificationTaskModule` (102) ArneBinder
* fix taskmodule metrics (`ignore_index`) (100) ArneBinder