Allow `utils.comm` to use custom backend

## 🚀 Feature

Third-party backends other than `torch.distributed` shall be allowed if users set it. This may enable training with torchelastic, apex or horovod.

## Motivation & Examples
```python
# my_comm_module.py:
# define a few necessary functions
def get_rank():
...
def ...

```

```python
detectron2.utils.comm.set_backend(my_comm_module)
# do other stuff as usual ..
```

## Note

Exact interface (e.g. what functions are required in custom backend) to be determined.

Start by reading the detectron2.utils.comm entry point and trace its current torch.distributed usage. Define which functions a custom backend must provide and how set_backend should affect existing communication calls; the issue is complete when a user-supplied backend can replace torch.distributed for the supported operations.

Written by the indexing model from the issue text.