Group fairness metrics can relabel missing group IDs

Nobody has claimed this yet.

When a batch does not contain every configured group, `BinaryGroupStatRates` currently compresses the group IDs.

```python
import torch
from torchmetrics.classification import BinaryGroupStatRates
from torchmetrics.functional.classification import binary_groups_stat_rates

preds = torch.tensor([1, 0])
target = torch.tensor([1, 0])
groups = torch.tensor([2, 2])

print(binary_groups_stat_rates(preds, target, groups, num_groups=3))
# {'group_0': tensor([0.5000, 0.0000, 0.5000, 0.0000])}

metric = BinaryGroupStatRates(num_groups=3)
metric.update(preds, target, groups)
print(metric.tp) # tensor([1, 0, 0])
```

The observations belong to group `2`, but they are returned and accumulated as group `0`. Negative and out-of-range IDs are similarly relabeled, and documented multidimensional group tensors fail in `bincount`.

This reproduces on current `master` (`8d008de1`) with Python 3.12.13 and PyTorch 2.13.0. I have a fix and regression tests ready :)

Start with BinaryGroupStatRates and binary_groups_stat_rates, reproducing the supplied example on the current master revision. Check the handling of missing, negative, out-of-range, and multidimensional group IDs, then run or extend the regression tests so group 2 remains group 2 and documented tensor shapes work without bincount errors.

Written by the indexing model from the issue text.