torch.save() fails when caching specific metrics due to lambda in preprocess

When caching metrics in `eval_multistep`, the process crashes for metrics that define a `preprocess` function as a local `lambda` (e.g. `PowerSpectrum`). Python cannot pickle lambda or locally defined functions, leading to a serialization failure during `torch.save`.

```
AttributeError: Can't get local object '<Class>.__init__.<locals>.<lambda>'
```