Support ReadConfig(input_context=) when number of shard is small

Currently in distributed setting, user can use tfds.ReadConfig(input_context=) as described in
https://www.tensorflow.org/datasets/performances#auto-shard_your_data_across_workers. This make sure that each worker read a different slice of the data.

However this only works when the number of shards is bigger than the number of workers. Otherwise, user is required to use the subsplit API:

def make_ds(input_context: tf.distribute.InputContext):
  split = tfds.even_splits('train', n=input_context.num_input_pipelines)[input_context.input_pipeline_id]
  ds = tfds.load('cifar10', split=split)
  return ds

ds = strategy.distribute_datasets_from_function(make_ds) 

It would be nice that the input_context automatically apply the subsplit API if info.splits[split].num_shards < read_config.input_context.num_input_pipelines, so the code in https://www.tensorflow.org/datasets/performances#auto-shard_your_data_across_workers would works for all datasets.

We might want to wait for the new TFDS file format before doing this.

Start with the ReadConfig(input_context=) auto-sharding behavior described in the TensorFlow Datasets performance documentation, then compare it with the even_splits example for datasets with too few shards. Done means the input_context path applies equivalent subsplitting when the shard count is smaller than the worker count, while preserving the existing behavior otherwise.