[Train] Per dataset execution_option for DataConfig

[xinyuangui2]

Summary

This PR adds support for per-dataset execution options in DataConfig, allowing users to specify different ExecutionOptions for different datasets. This enables fine-grained control over how each dataset is processed by Ray Data.

Use case

For train and test dataset, users might need different execution_options.exclude_resources or execution_options.resource_limits. This PR enables this configuration.

For example, it's common to have different ExecutionOptions for train dataset and validation dataset.

trainer = ray.train.torch.TorchTrainer(
    train_func,
    # Pass training dataset in datasets arg to split it across training workers
    datasets={"train": train_dataset, "val": val_dataset},
    scaling_config=ray.train.ScalingConfig(
        num_workers=2,
        use_gpu=True,
        # Use powerful GPUs for training
        accelerator_type="A100",
    ),
    dataset_config=DatasetConfig(
        datasets_to_split='all',
        execution_options={
            # train dataset uses default execution options
            'val': ExecutionOptions(exclude_resources=ExecutionResources.zero(), resource_limits=ExecutionResources.for_limits(cpu=1)),
        }
        enable_shard_locality=True,
    ),
)

[justinvyu]

great, thanks! can you add a motivating usage example in the PR description?

[justinvyu]

Thanks! Almost good

[justinvyu]

Thanks!

[justinvyu]

Btw, can you add a motivating usage example in the PR description?

[xinyuangui2]

Btw, can you add a motivating usage example in the PR description?

Done.