bugfix - runtime error "sharding is ambiguous"

When trying to load the dataset with the split "microphone-single", I
get following error:
```
RuntimeError: Sharding is ambiguous for this dataset: we found several data sources lists of different lengths, and we don't know over which list we should parallelize:
- key samples_paths has length 134
- key ids has length 136
- key verification_ids has length 134
To fix this, check the 'gen_kwargs' and make sure to use lists only for data sources, and use tuples otherwise. In the end there should only be one single list, or several lists with the same length.
```

This is due to `ids` still containing the missing files for that split.

ami.py

@@ -371,10 +371,10 @@ class AMI(datasets.GeneratorBasedBuilder):
         annotation_path = dl_manager.download_and_extract(_DL_URL_ANNOTATIONS)
 
         # train
-        train_files = [path.format(_id, _id) for _id in _TRAIN_SAMPLE_IDS for path in self.config.dl_path_formats]
-        train_files = list(
-            filter(lambda f: f.split("/")[-1].split(".")[0] not in self.config.missing_files, train_files)
+        train_sample_ids = list(
+            filter(lambda f: f not in self.config.missing_files, _TRAIN_SAMPLE_IDS)
         )
+        train_files = [path.format(_id, _id) for _id in train_sample_ids for path in self.config.dl_path_formats]
         train_ids = [f.split("/")[-1].split(".")[0] for f in train_files]
         train_path = dl_manager.download_and_extract(train_files)
 
@@ -396,7 +396,7 @@ class AMI(datasets.GeneratorBasedBuilder):
                 gen_kwargs={
                     "annotation_path": annotation_path,
                     "samples_paths": train_path,
-                    "ids": _TRAIN_SAMPLE_IDS,
+                    "ids": train_sample_ids,
                     "verification_ids": train_ids,
                 },
             ),