Flatten for eigen methods

When using eigsh routines from scipy or cupy, we have to output a flat tensor from the matvec function. However, when calling the hvp on the functional call to the model, the output is a dict of tensors. Right now, we have to flatten after each hvp call, e.g. for one batch:

```python
def batch_hvp(vec: torch.Tensor):
return flatten_tensors_to_vector(
hvp(
batch_loss,
model_params,
align_structure(model_params, vec),
reverse_only=reverse_only,
).values()
)
```

We should investigate the runtime and memory consequences of this and if there is an alternative for it.