This PR updates the chunk iterator to make it more general for different tasks.

I added a new argument excluded_key_prefixes for ChunkIterFactory to allow certain keys to be ignored when checking the length consistency for each sample:

            for key in sequence_keys:
-               if len(batch[key]) != len(
-                   batch[sequence_keys[0]]
-               ) and not key.startswith("enroll_ref"):
+               if key.startswith(self.excluded_key_prefixes):
+                   # ignore length inconsistency for `excluded_key_prefixes`
+                   continue
+               if len(batch[key]) != len(batch[sequence_keys[0]]):
                    raise RuntimeError(
                        f"All sequences must has same length: "
                        f"{len(batch[key])} != {len(batch[sequence_keys[0]])}"

This can be useful for audio-to-audio tasks like target speaker extraction (TSE), where additional features are required as input, which do not necessarily have the same length as the input/target signal.

In our previous discussion in the SE meeting, we decided to make a new chunk iterator for such purpose. But after some coding, I find the current minimal changes in espnet2/iterators/chunk_iter_factory.py are not so bad.

If you still feel it necessary to make a new script, I will do it.