Functional nominal metrics raise RuntimeError when category labels do not start at 0

Nobody has claimed this yet.

## 🐛 Bug

Functional nominal metrics (`cramers_v`, `pearsons_contingency_coefficient`, `theils_u`, `tschuprows_t`) raise a `RuntimeError` during confusion matrix calculation when categorical inputs contain non-zero-based or non-contiguous category labels (e.g., labels `[1, 2]` or `[5, 6]`).

### To Reproduce

Code sample

```python
import torch
from torchmetrics.functional.nominal import cramers_v

preds = torch.tensor([5, 6, 5, 6, 5, 6, 5, 6])
target = torch.tensor([5, 6, 5, 5, 6, 6, 5, 6])

# Raises RuntimeError: shape '[2, 2]' is invalid for input of size 19
cramers_v(preds, target)
```

Environment

- TorchMetrics version: 1.9.0 (latest master)
- Python & PyTorch Version: Python 3.10+, PyTorch 2.x
- Any other relevant information such as OS: All platforms (Linux, macOS, Windows)

### Additional context

The functional API infers `num_classes = len(torch.cat([preds, target]).unique())` but passes the unmapped raw values to `_multiclass_confusion_matrix_update`, which expects values strictly in `[0, num_classes - 1]`. Any values outside this range cause invalid bin indexing and reshape failures. Additionally, `_compute_chi_squared` performs in-place float addition on integer `confmat` tensors, causing type cast issues. Remapping categorical labels to contiguous 0-based indices before updating the confusion matrix resolves this issue.

Start with the functional nominal metric entry points for cramers_v, pearsons_contingency_coefficient, theils_u, and tschuprows_t, then trace their use of _multiclass_confusion_matrix_update and _compute_chi_squared. Reproduce the supplied labels [5, 6] and verify that all four metrics calculate successfully without reshape or tensor type-cast errors.

Written by the indexing model from the issue text.