Improve dataset `.skip()` speed in streaming mode

### Feature request

Add extra information to the `dataset_infos.json` file to include the number of samples/examples in each shard, for example in a new field `num_examples` alongside `num_bytes`. The `.skip()` function could use this information to ignore the download of a shard when in streaming mode, which AFAICT it should speed up the skipping process.

### Motivation

When resuming from a checkpoint after a crashed run, using `dataset.skip()` is very convenient to recover the exact state of the data and to not train again over the same examples (assuming same seed, no shuffling). However, I have noticed that for audio datasets in streaming mode this is very costly in terms of time, as shards need to be downloaded every time before skipping the right number of examples.

### Your contribution

I took a look already at the code, but it seems a change like this is way deeper than I am able to manage, as it touches the library in several parts. I could give it a try but might need some guidance on the internals.