Datapipe handling with more than 1 GPU gives torch.multiprocessing.spawn.ProcessExitedException: process 1 terminated with exit code 1

Nobody has claimed this yet.

## 🐛 Bug

A dataloader around a datapipe works in cpu and gpu but not with 2 gpus

### To Reproduce

Code to reproduce (you need 2 gpus)
https://gist.github.com/franchesoni/7c048c0f0eb56f6aaa2731b0ead78a51#file-issue-py

Output (logged by ussing `python issue.py >> issue.out 2>&1`
https://gist.github.com/franchesoni/7c048c0f0eb56f6aaa2731b0ead78a51#file-issue-out

### Expected behavior

Not break

### Environment

* CUDA:
- GPU:
- TITAN V
- GeForce RTX 2080 Ti
- available: True
- version: 11.3
* Packages:
- numpy: 1.22.3
- pyTorch_debug: False
- pyTorch_version: 1.11.0
- pytorch-lightning: 1.6.3
- tqdm: 4.64.0
* System:
- OS: Linux
- architecture:
- 64bit
- ELF
- processor: x86_64
- python: 3.9.12
- version: #80~18.04.1-Ubuntu SMP Mon Apr 12 23:26:25 UTC 2021
### Additional context

cc @otaj @justusschock @awaelchli @ninginthecloud @rohitgr7 @kaushikb11 @akihironitta

Start with the linked issue.py reproduction and issue.out output, running it first with one GPU and then with two GPUs. Trace the dataloader and datapipe path around the reported torch.multiprocessing.spawn.ProcessExitedException. Done means the datapipe-backed dataloader runs on two GPUs without the process exiting with code 1.

Written by the indexing model from the issue text.