Make MLuke tokenizer tests slow

[sgugger]

What does this PR do?

To avoid the timeouts in the test suites, this PR makes all mLuke tokenizer tests slow until we figure out a way to speed them up.

[LysandreJik]

cc @NielsRogge

[ryokan0123]

Hi, I suspect that the slowness is from reading the large entity_vocab.json in MLukeTokenizer.__init__(), which has 418 MB (studio-ousia/mluke-base).
When I tried on my local PC, only reading the entity vocab file in __init__() took around 20s.

If this is the problem, we can reduce the file size.
The entity vocabulary has 120M entries, but now we naively represent the same entities from different languages separate entries, like

{
	“en:Japan”: 13,
	“de:Japan”: 13,
	'ko:일본’: 13,
	“ja:日本”: 13,
	...
}

We could instead represent the entities with Wikidata QID (e.g., Q17 for Japan), which would significantly reduce the file size. Using QID may look cryptic at first, but now we think this may be the right way to handle entities in multilingual settings.
Do you have any suggestions on this?

[sgugger]

For tests we don't need a real file (apart from the integration test). All other tokenizers are tested on a vocab of 10-15 tokens, so ideally building something that just have the strict minimum to test would be great.

[ryokan0123]

Sure, I will work on it.