add a test to match author's tokenization

[SaulLu]

What does this PR do?

I propose to add a test that aims to verify that we are able to reproduce the tokenization produced by the authors of MarkUpLM on the downstream task of WebSRC.

To design this test, I therefore ran the authors' code on the WebSRC downstream task and isolated:

• the xpaths in the get_xpath_and_treeid4tokens method

https://github.com/microsoft/unilm/blob/e4929f812398207b7fefb4dda6e9debcb8ce34b9/markuplm/examples/fine_tuning/run_websrc/utils.py#L256-L284

• the other targets in the feature creation stage (in the convert_examples_to_features method):

https://github.com/microsoft/unilm/blob/e4929f812398207b7fefb4dda6e9debcb8ce34b9/markuplm/examples/fine_tuning/run_websrc/utils.py#L664-L684

This test already reveals some differences:

• On the slow tokenizer the template on the phrase pair is not the same. The authors have <s> question </s> content </s> and we have <s> question </s></s> content .
• The slow tokenizer loses the last tokens and replaces them with pad tokens
• The fast tokenizer distorts the tokenization of the question

[SaulLu]

@NielsRogge , I can't add you as a reviewer to this PR 🙂 . I would love to have your review on it