ModelCheckpoint tries to find monitored key on non-zero ranks

Nobody has claimed this yet.

### Bug description

Hi, we're using `ModelCheckpoint` with `save_top_k=2` to keep track of the best checkpoints based on `monitor=val_auc`.

We're computing and logging `val_auc` on `validation_epoch_end`, however we only do it on *rank zero*.

```python
def validation_epoch_end(self, outputs):
gathered_outputs = gloo_gather(outputs)
if rank_zero():
auc = compute_auc(gathered_outputs)
self.log('val_auc': auc)
```

If we run this model over >1 GPUs, we get an error:
```
pytorch_lightning.utilities.exceptions.MisconfigurationException: `ModelCheckpoint(monitor='val_auc')` could not find the monitored key in the returned metrics: ['train_loss', 'val_loss', 'epoch', 'step']. HINT: Did you call `log('val_auc, value)` in the `LightningModule`?
```

This tells us that `ModelCheckpoint` ran on other ranks too, not just on rank zero. This looks like a wrong behaviour - model checkpoint shouldn't run on non-zero ranks, in our understanding.

Is our configuration wrong or is this a bug? Thanks.

Related error was posted already in discussions but not answered: https://github.com/Lightning-AI/lightning/discussions/14806

### How to reproduce the bug

```python
Log a metric only on rank zero and run training with `ModelCheckpoint` callback and `save_top_k=2`.
```

### Error messages and logs

```
pytorch_lightning.utilities.exceptions.MisconfigurationException: `ModelCheckpoint(monitor='val_auc')` could not find the monitored key in the returned metrics: ['train_loss', 'val_loss', 'epoch', 'step']. HINT: Did you call `log('val_auc, value)` in the `LightningModule`?
```

### Environment

```
#- PyTorch Lightning Version (e.g., 1.5.0): 1.7.7
#- PyTorch Version (e.g., 1.10): 1.12.1+cu116
#- Python version (e.g., 3.9): 3.8.10
#- OS (e.g., Linux): Linux
#- How you installed Lightning(`conda`, `pip`, source): pip
```

### More info

_No response_

cc @borda @awaelchli @carmocca @Blaizzy

Start at the ModelCheckpoint callback entry point and reproduce the reported multi-GPU case with save_top_k=2, monitor=val_auc, and the metric logged only on rank zero. Trace how the callback handles monitored metrics on non-zero ranks. Done means the reproduction no longer raises the missing-key MisconfigurationException while checkpoint monitoring still works.

Written by the indexing model from the issue text.