feat(framework) Use dataset caching in flwr new templates

[jafermarq]

Caching the FederatedDataset object creation and its partitioner makes simulations more efficient.

Additional changes:

• use full names of the dataset e.g. instead of mnist use ylecun/mnist (related to recent names change on HF Hub; both work, though mnist can't be found right now, it's available under the ylecun/mnist name)
• trust_remote_code=True in FederatedDataset for mnist dataset to avoid the warning of the dataset that does NOT have file representation on HF Hub (note the remaining dataset in the templates have file representation on HF Hub)
• explicit import of IidPartitioner to simplify user to switch to other partitioning schemes (instead of using the partitioners={split-name: num_partitions} shorter API now it's partitioners={split-name: IidPartitioner(num_partitions)}