Xarray version 2026.4.0 breaks GraphCast Demo Notebook: Passing a Dataset to the constructor is not supported

There's [yet another](https://github.com/google-deepmind/graphcast/issues/212) xarray 2026.4.0 incompatibility...

# Problem

These lines in [`graphcast/rollout.py` line 320](https://github.com/google-deepmind/graphcast/blob/main/graphcast/rollout.py#L310-L313):
```python
# Create copies to avoid mutating inputs.
inputs = xarray.Dataset(inputs)
targets_template = xarray.Dataset(targets_template)
forcings = xarray.Dataset(forcings)
```
produce following error when running the rollout cell of the notebook cell inside `graphcast_demo.ipynb`:
> TypeError: Passing a Dataset as `data_vars` to the Dataset constructor is not supported. Use `ds.copy()` to create a copy of a Dataset.

# Reason

Again, the reason is [breaking changes in Xarray 2026.4.0](https://docs.xarray.dev/en/latest/whats-new.html#v2026-04-0-apr-13-2026):
> Passing a Dataset as data_vars to the Dataset constructor now raises TypeError. This was never intended behavior and silently dropped attrs. Use [Dataset.copy()](https://docs.xarray.dev/en/latest/generated/xarray.Dataset.copy.html#xarray.Dataset.copy) instead ([GH11095](https://github.com/pydata/xarray/issues/11095)). By [Kristian Kollsga](https://github.com/kkollsga).

# Proposed Solution

```python
# Create copies to avoid mutating inputs.
inputs = inputs.copy()
targets_template = targets_template.copy()
forcings = forcings.copy()
```

(I did not check the complete repository for further deprecated usages!)

Start with graphcast/rollout.py around line 320 and inspect the rollout cell in graphcast_demo.ipynb. Run the notebook's rollout cell with xarray 2026.4.0 and verify that the Dataset copies no longer raise the reported TypeError and the demo continues successfully.

Written by the indexing model from the issue text.