Describe the bug
The model cannot be saved in ONNX format as shown in the flair/resources/docs/TUTORIAL_13_TRANSFORMERS_PRODUCTION.md tutorial. The export raises the following error:
RuntimeError: Dynamic shape axis should be no more than the shape dimension for sequ_length

To Reproduce

from flair.data import Sentence
from flair.models import SequenceTagger
from flair.embeddings import TransformerWordEmbeddings, TransformerDocumentEmbeddings
import flair.datasets

if __name__ == '__main__':
    model = SequenceTagger.load("ner-large")
    assert isinstance(model.embeddings, (TransformerWordEmbeddings, TransformerDocumentEmbeddings))

    # Serialize the model (use sample sentences)
    sentences = list(flair.datasets.CONLL_03_SPANISH().test)[:5]
    model.embeddings = model.embeddings.export_onnx("flert-embeddings.onnx", sentences, providers=["CUDAExecutionProvider", "CPUExecutionProvider"])

Expected behavior
Model saved in ONNX format

Environment (please complete the following information):

• OS: Linux
• Flair version: master (last commit: ee7e619)
• ONNX runtime version: 1.12.0

Additional context
I found out that this problem arises when the torch.onnx.export is called. While debugging, I found out that this comes from the order of example_tensors. The model expects the following order ['input_ids', 'lengths', 'attention_mask', 'overflow_to_sample_mapping', 'word_ids'] so I suggest that the example_tensors is reordered accordingly (quick fix):

desired_order = ['input_ids', 'lengths', 'attention_mask', 'overflow_to_sample_mapping', 'word_ids']
example_tensors = {k: example_tensors[k] for k in desired_order}

And while we are at it, I would suggest to fix the expected type of the args argument of torch.onnx.export (should be a tuple)

        torch.onnx.export(
            embedding,
            tuple(example_tensors.values()),
            ...
        )