Fix dataset formats to match each pipeline expectations

[russellb]

generate_data.py calls leaf_node_to_samples(), which right now assumes that the expected dataset format is the same for all pipeline types. It matched the simple pipelines and the full knowledge pipeline. I didn't notice that the full skills pipeline was different. I saw that #50 changed the name of fields expected in the template, so now we have 3 different dataset formats expected by the default pipelines.

simple pipelines expect question_N, answer_N for N in (1, 2, 3)
full knowledge pipeline now expects icl_query_N / icl_reponse_Nfor N in (1, 2, 3)
full skills pipelines expect seed_question and seed_response (just 1 instead of 3)

Short term fix:

1. Document the dataset format expected by each pipeline.
2. fix the code to generate the dataset in the expected format for each pipeline.

Longer term consideration:

None of this matches the source format for the dataset (taxonomy). Allowing people to specify a custom pipeline implies specifying their expected sample dataset format somehow.

Another idea instead ...

1. Always assume a consistent dataset format.
2. Add a new pipeline capability for dataset transformation -- rename fields if you want, squash the rows into groups of 3 seed questions/answers per row (for the knowledge case)

I think something like this is going to be necessary to allow more configurable custom pipelines, as we'll need a way for a custom pipeline to declare the dataset format it is expecting from a known starting point.