refactor code to calculate records per shard using n_volumes and number of shards

[hvgazula]

nobrainer/nobrainer/dataset.py

Lines 115 to 122 in 976691d

	first_shard = (
	dataset.take(1)
	.flat_map(
	lambda x: tf.data.TFRecordDataset(x, compression_type=compression_type)
	)
	.map(map_func=parse_fn, num_parallel_calls=num_parallel_calls)
	)
	block_length = len([0 for _ in first_shard])

If the number of volumes in the shard is too large, this snippet of code can be time-consuming. Alternatives are

• use a combination of n_volumes and number of files with file_pattern to calculate len(first_shard)
• provide metadata (number of volumes in the shard) as well as total number of volumes in the dataset

[hvgazula]

code for option 1: https://github.com/neuronets/nobrainer_training_scripts/blob/784c8668ae01356173faffbcf860bca458f46a73/1.2.0/create_tfshards.py#L307-L311 should now work after #329

[hvgazula]

Ideally, if the tfrecords are created using the API, with the aforementioned change, we can ensure the same number of records in every shard except the last one. Now, if n_volumes is not specified, it can be calculated using this function, which is num_records_first_shard * (num_shards - 1) + num_records_in_last_shard