gh-3097: fix multitask model training

[helpmefindaname]

This PR fixes #3097 by introducing an evaluate_all parameter. That way, users can use multitask_model.evaluate(corpus.test, gold_label_type="Task_0", evaluate_all=False) to evaluate the first task. If evaluate_all is set to true, the value of gold_label_type will be ignored and every task will be evaluated on their respective label type.

While doing this implementation I also noticed, that the evaluation currently evaluates sentences that have multiple tasks annotated to be evaluated for only one random task. -> now the evaluation uses all sentences for all tasks they are assigned to as it should be expected.

I also added the possibilities to compute the loss of a sentence on all assigned tasks during training. Hence Training Knowledge Graph Construction (NER + Relation Extraction + NEL) with a shared transformer embedding, would lead to only slightly increased training time compared to NER only, while leaveraging information of all 3 tasks at any time.

I found a bug where the embeddigns where not embedded right, as the identify_dynamic_embeddings only looks at the first sentence of the batch which could be unenbedded (e.g. for relation extraction or NEL but has no NER labels). Now the logic searches the whole batch and continues searching until it finds a batch that contains ANY embedding.

[alanakbik]

@helpmefindaname thanks a lot for fixing and improving this!