Support parallelized downloading and processing in load_dataset with Spark

### Feature request

When calling `load_dataset` for datasets that have multiple files, support using Spark to distribute the downloading and processing job to worker nodes when `cache_dir` is a cloud file system shared among nodes.

```python
load_dataset(..., use_spark=True)
```

### Motivation

Further speed up `dl_manager.download` and `_prepare_split` by distributing the workloads to worker nodes.

### Your contribution

I can submit a PR to support this.