Merge DatasetDict and Dataset

As discussed in #2424 and #2437 (please see there for detailed conversation):
- It would be desirable to improve UX with respect the confusion between DatasetDict and Dataset.
- The difference between Dataset and DatasetDict is an additional abstraction complexity that confuses "typical" end users.
- A user expects a "Dataset" (whatever it contains multiple or a single split) and maybe it could be interesting to try to simplify the user-facing API as much as possible to hide this complexity from the end user.

Here is a proposal for discussion and refined (and potential abandon if it's not good enough):
- let's consider that a DatasetDict is also a Dataset with the various split concatenated one after the other
- let's disallow the use of integers in split names (probably not a very big breaking change)
- when you index with integers you access the examples progressively in split after the other is finished (in a deterministic order)
- when you index with strings/split name you have the same behavior as now (full backward compat)
- let's then also have all the methods of a Dataset on the DatasetDict

The end goal would be to merge both Dataset and DatasetDict object in a single object that would be (pretty much totally) backward compatible with both.

There are a few things that we could discuss if we want to merge Dataset and DatasetDict:

1. what happens if you index by a string ? Does it return the column or the split ? We could disallow conflicts between column names and split names to avoid ambiguities. It can be surprising to be able to get a column or a split using the same indexing feature
```
from datasets import load_dataset

dataset = load_dataset(...)
dataset["train"]
dataset["input_ids"]
```
2. what happens when you iterate over the object ? I guess it should iterate over the examples as a Dataset object, but a DatasetDict used to iterate over the splits as they are the dictionary keys. This is a breaking change that we can discuss.

Moreover regarding your points:

- integers are not allowed as split names already
- it's definitely doable to have all the methods. Maybe some of them like train_test_split that is currently only available for Dataset can be tweaked to work for a split dataset

cc: @thomwolf @lhoestq