Optimise behaviour in respect to indices mapping

_Originally [posted](https://huggingface.slack.com/archives/C02V51Q3800/p1675443873878489?thread_ts=1675418893.373479&cid=C02V51Q3800) on Slack_

Considering all this, perhaps for Datasets 3.0, we can do the following:
* [ ] have `continuous=True` by default in `.shard` (requested in the survey and makes more sense for us since it doesn't create an indices mapping)
* [x] allow calling `save_to_disk` on "unflattened" datasets
* [ ] remove "hidden" expensive calls in `save_to_disk`, `unique`, `concatenate_datasets`, etc. For instance, instead of silently calling `flatten_indices` where it's needed, it's probably better to be explicit (considering how expensive these ops can be) and raise an error instead