Current implementation of `DistributedSampler` is ideal for distributed
training using map datasets, as they fit in memory and have known size.
However, it doesn't support distributed training using `IterableDataset`
datasets, as these classes do not implement `__len__`.
To fix that, a `length` keyword was added to `DistributedSampler`, which
has precedence when set.

An extra `padding=True` parameter was also added was give finer control
on whether the (returned) index list should be padded by the sampler.
This is useful for preventing duplicate reading on `IterableDataset`
datasets that do not fit in memory or which data reading or transformation
are expensive.

Finally, set_rank method was added, similarly the existing `set_epoch`,
to ease distributed training. When `DataLoader` is created with
`num_workers` > 0 and `dataset` is an instance of `ChunkDataset`,
a copy of `DistributedSampler` on each worker needs to be configured
with their new rank.

There is no back compatibility with this change.