Updated RACE to use datasets instead of nlp package.

lm_eval/tasks/race.py

 from . common import HFTask
 from ..utils_stream import X, each, apply, join, filt, one
 import collections
-import nlp
+import datasets
 
 
 class RACE(HFTask):
@@ -26,7 +26,7 @@ class RACE(HFTask):
         # is shown that one document is made per passage.
 
         r = collections.defaultdict(list)
-        for item in nlp.load_dataset(path=self.DATASET_PATH, name=self.DATASET_NAME)[set]:
+        for item in datasets.load_dataset(path=self.DATASET_PATH, name=self.DATASET_NAME)[set]:
             r[item['article']].append(item)
         
         res = list(r.values() >> each(lambda x: {