Implement `UnzipperIterDataPipe`

🚀 The feature

Add unzip as requested by both Vision and Text teams.

dp = IterableWrapper([(1, 2, 3), (4, 5, 6), (7, 8, 9)])
dp0, dp1, dp2 = dp.unzip()

list(dp0)  # [1, 4, 7]

We may be able to reuse the logic of _ChildDataPipehttps://github.com/pytorch/pytorch/blob/fa38e93fe98bfcbdb11288ecfbe5af5c264aabb3/torch/utils/data/datapipes/iter/combining.py#L145

Support tuple, list maybe dict?

Motivation, pitch

It’s way better than letting users to use the following script to mimic unzip

dp = IterableWrapper([(1, 2, 3), (4, 5, 6), (7, 8, 9)])
def expand_fn(data):
    index_data = []
    for i, d in enumerate(data):
        index_data.append((i, d))
    return index_data

dp = dp.flatmap(expand_fn)

def classify(data):
    return data[0]

dp0, dp1, dp2 = dp.demux(classify)

...

Alternatives

No response

Additional context

Note: We need to take extra care for these demux, fork, and future unzip. If users do dp0, dp1, _ = dp.unzip(), this pipeline may be unserializable after iteration as the third output DataPipe is never executed.

[erip]

If users do dp0, dp1, _ = dp.unzip()

Does it make sense to add a ignore_idx: Union[int, Tuple[int, ...]] which allows unzip to skip over certain elements within the inner dp? This avoids the problem, so you’d have something like

dp0, dp1 = dp.unzip(ignore_idx=2) # or dp.unzip(ignore_idx=(2,)), dp.unzip(ignore_idx=(-1,)), ... 

[NivekT]

Closed by #198