`load_dataset` has ~4 seconds of overhead for cached data

### Feature request

When loading a dataset that has been cached locally, the `load_dataset` function takes a lot longer than it should take to fetch the dataset from disk (or memory).
This is particularly noticeable for smaller datasets. For example, wikitext-2, comparing `load_data` (once cached) and `load_from_disk`, the `load_dataset` method takes 40 times longer.

⏱ 4.84s ⮜ load_dataset
⏱ 119ms ⮜ load_from_disk

### Motivation

I assume this is doing something like checking for a newer version.

If so, that's an age old problem: do you make the user wait _every single time they load from cache_ or do you do something like load from cache always, _then_ check for a newer version and alert if they have stale data. The decision usually revolves around what percentage of the time the data will have been updated, and how dangerous old data is.

For most datasets it's extremely unlikely that there will be a newer version on any given run, so 99% of the time this is just wasted time.

Maybe you don't want to make that decision for all users, but at least having the _option_ to not wait for checks would be an improvement.

### Your contribution

.