Checking that split name is correct happens only after the data is downloaded

### Describe the bug

Verification of split names (=indexing data by split) happens after downloading the data. So when the split name is incorrect, users learn about that only after the data is fully downloaded, for large datasets it might take a lot of time.

### Steps to reproduce the bug

Load any dataset with random split name, for example:
```python
from datasets import load_dataset

load_dataset("mozilla-foundation/common_voice_11_0", "en", split="blabla")
```
and the download will start smoothly, despite there is no split named "blabla".

### Expected behavior

Raise error when split name is incorrect.

### Environment info

`datasets==2.9.1.dev0`