Xarray version 2026.4.0 breaks GraphCast Demo Notebook: "time" coord parsed as `int32` instead `timedelta64`

The new xarray version parses NetCDF dataset "time" coord as "int32" instead "timedelta". The issue occurs when parsing the NetCDF example weather data files loaded from google cloud bucket.

## Problem
Inside the `graphcast_demo.ipynb`, section "Load the example data", a NetCDF file is loaded and opened with xarray:
```python
with gcs_bucket.blob(f"{dir_prefix}dataset/{dataset_file.value}").open("rb") as f:
example_batch = xarray.load_dataset(f).compute()
```
The new xarray version parses the `time` coordinate of the dataset as `int32`. But the graphcast code excpects "An xarray.Dataset with a 'time' dimension whose coordinates are timedeltas.[…]" (graphcast.data_utils.extract_input_target_times).
This leads to the error in `graphcast/graphcast/data_utils.py:282`
> TypeError: Concatenation operation is not implemented for NumPy arrays, use np.concatenate() instead. Please do not rely on this error; it may not be given on all Python implementations.

## Reason

Xarray version [2026.4.0 finalizes the deprecation timedelta decoding via units](https://github.com/pydata/xarray/releases/tag/v2026.04.0):
> Xarray will now no longer by default decode variables with only a timedelta-like "units" attribute into np.timedelta64 values. […]
> https://github.com/pydata/xarray/pull/11173

This means that the "time" coordinate of NetCDF files like `graphcast/dataset/source-era5_date-2022-01-01_res-0.25_levels-37_steps-01.nc` will not be parsed as `timedelta64` even though it has the attribute `units: hours`.

## Proposed Solution
Enforce previous behaviour by explicitly setting `decode_timedelta=True`:

```python
with gcs_bucket.blob(f"{dir_prefix}dataset/{dataset_file.value}").open("rb") as f:
example_batch = xarray.load_dataset(f, decode_timedelta=True).compute()
```

Start in graphcast_demo.ipynb at the “Load the example data” section and inspect the xarray.load_dataset call for the NetCDF file. Run that notebook path with the example weather data; done means the time coordinate is decoded as timedeltas and the later GraphCast data processing no longer raises the reported error.

Written by the indexing model from the issue text.