Fixes scrolls task bug with few_shot examples (#2003)

Bug:

```
python -m scripts.write_out --task scrolls_quality --output_base_path ~/workspace/
Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/lm-evaluation-harness/scripts/write_out.py", line 92, in <module>
    main()
  File "/lm-evaluation-harness/scripts/write_out.py", line 51, in main
    task_dict = tasks.get_task_dict(task_names, task_manager)
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 423, in get_task_dict
    task_name_from_string_dict = task_manager.load_task_or_group(
                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 271, in load_task_or_group
    collections.ChainMap(*map(self._load_individual_task_or_group, task_list))
  File "/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 162, in _load_individual_task_or_group
    return load_task(task_config, task=name_or_config, group=parent_name)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 148, in load_task
    task_object = config["class"]()
                  ^^^^^^^^^^^^^^^^^
  File "/lm-evaluation-harness/lm_eval/tasks/scrolls/task.py", line 120, in __init__
    super().__init__()
  File "/lm-evaluation-harness/lm_eval/api/task.py", line 703, in __init__
    self._config = TaskConfig(**config)
                   ^^^^^^^^^^^^^^^^^^^^
TypeError: lm_eval.api.task.TaskConfig() argument after ** must be a mapping, not NoneType
```

lm_eval/tasks/scrolls/task.py

@@ -4,12 +4,12 @@ from functools import reduce
 
 import numpy as np
 import transformers.data.metrics.squad_metrics as squad_metrics
-from datasets import load_metric
+from datasets import Dataset, load_metric
 from transformers import AutoTokenizer
 
 from lm_eval.api.instance import Instance
 from lm_eval.api.metrics import mean
-from lm_eval.api.task import Task
+from lm_eval.api.task import ConfigurableTask
 
 
 _CITATION = """
@@ -108,7 +108,7 @@ def _num_cpu_cores():
         return len(os.sched_getaffinity(0))
 
 
-class _SCROLLSTask(Task):
+class _SCROLLSTask(ConfigurableTask):
     VERSION = 2
     DATASET_PATH = "tau/scrolls"
     DATASET_NAME = None
@@ -117,7 +117,7 @@ class _SCROLLSTask(Task):
     PRUNE_NUM_PROC = None
 
     def __init__(self):
-        super().__init__()
+        super().__init__(config={"metadata": {"version": self.VERSION}})
         if self.DATASET_NAME is not None:
             self.metric = load_metric(_download_metric(), config_name=self.DATASET_NAME)
 
@@ -131,12 +131,26 @@ class _SCROLLSTask(Task):
         return False
 
     def training_docs(self):
-        for doc in self.dataset["train"]:
-            yield from self._process_doc(doc)
+        processed_docs = list(map(self._process_doc, self.dataset["train"]))
+        
+        # Flatten the list of lists since _process_doc returns a list of one element.
+        processed_docs = [item for sublist in processed_docs for item in sublist]
+        processed_dict = {
+            key: [d[key] for d in processed_docs] for key in processed_docs[0]
+        }
+
+        return Dataset.from_dict(processed_dict)
 
     def validation_docs(self):
-        for doc in self.dataset["validation"]:
-            yield from self._process_doc(doc)
+        processed_docs = list(map(self._process_doc, self.dataset["validation"]))
+        
+        # Flatten the list of lists since _process_doc returns a list of one element.
+        processed_docs = [item for sublist in processed_docs for item in sublist]
+        processed_dict = {
+            key: [d[key] for d in processed_docs] for key in processed_docs[0]
+        }
+        
+        return Dataset.from_dict(processed_dict)
 
     def should_decontaminate(self):
         return True