split cache used when processing different split

## Describe the bug`
```
ds1 = load_dataset('squad', split='validation')
ds2 = load_dataset('squad', split='train')
ds1 = ds1.map(some_function)
ds2 = ds2.map(some_function)
assert ds1 == ds2
```
This happens when ds1 and ds2 are created in `pytorch_lightning.DataModule` through

```
class myDataModule:

def train_dataloader(self):
ds = load_dataset('squad', split='train')
ds = ds.map(some_function)
return [ds]

def val_dataloader(self):
ds = load_dataset('squad', split="validation")
ds = ds.map(some_function)
return [ds]
```
I don't know if it depends on `pytorch_lightning` or `datasets` but setting `ds.map(some_function, load_from_cache_file=False)` fixes the issue.

If this is not enough to replicate I will try and provide and MWE, I don't have time now so I thought I wuld open the issue first!