Dataset index and slicing

[davidslater]

Fixes #871

This PR does a few related things:

1. Renames "split_type" to "split" to make it clear that it is verbatim what is given to TFDS load (same kwarg name)
   https://www.tensorflow.org/datasets/api_docs/python/tfds/load

2. Modifies scenarios so that configs can override the train and eval splits. This enables TFDS slicing operations on them such as "test[:10]" or "train_clean100+train_clean360+train_other500":
   https://www.tensorflow.org/datasets/splits#examples

3. Adds additional parsing of splits to enable individual indexing for a specific data point, such as "test[355]", and indexing based off of a list, "test[[1, 4, 5, 7, 9]]". Under the hood, these are translated to single-element slices. Advanced slicing operations involving a step parameter, such as "test[10:20:2]" are not permitted, and raise a NotImplementedError.

[davidslater]

Biggest code changes are in datasets.py and its associated test file.

[lcadalzo]

What's the intended behavior when overlapping splits are provided? E.g. "train[1:3] + train[2]". The current behavior is to repeat elements, so in this case train[2] is returned in two consecutive batches.

[lcadalzo]

Is it of interest to allow for specifying negative indices when (1) specifying a single index and/or (2) when providing a list of indices?

(1) E.g, if I only want to evaluate on the penultimate sample, specifying "train[-2]" will cause an error. "train[-2:-1]", however, will yield the one sample I want.

(2) E.g., if I want to evaluate on the first and penultimate sample, "train[[0, -2]]" will cause an error since only positive indices are supported when passing a list. And "train[0] + train[-2:-1]" would work.

[davidslater]

What's the intended behavior when overlapping splits are provided? E.g. "train[1:3] + train[2]". The current behavior is to repeat elements, so in this case train[2] is returned in two consecutive batches.

The intended behavior is simply to match the underlying TFDS behavior, which allows duplicate and out-of-order splits. I don't expect people to make much use of it, but it is much simpler on our end than trying to determine uniqueness.

[davidslater]

Is it of interest to allow for specifying negative indices when (1) specifying a single index and/or (2) when providing a list of indices?

(1) E.g, if I only want to evaluate on the penultimate sample, specifying "train[-2]" will cause an error. "train[-2:-1]", however, will yield the one sample I want.

(2) E.g., if I want to evaluate on the first and penultimate sample, "train[[0, -2]]" will cause an error since only positive indices are supported when passing a list. And "train[0] + train[-2:-1]" would work.

I had not considered using negative indices. I didn't realize they would work. I can make an update to allow them if you think that makes sense.

[lcadalzo]

I think, since it's unlikely they would be commonly used, we can hold off and only add later if it's something that a user is specifically looking for.

[lcadalzo]

LGTM