Refactor AzureML kNN pipeline

[riedgar-ms]

Refactor the kNN pipeline so that it is a separately callable thing. This constructs a subpipeline which:

• Accepts a guidance program and two datasets:
  • The 'input_dataset of questions to be answered
  • The examples_dataset of questions to be used as few-shot examples
  • Validates the schema of the two input datasets
  • Computes embeddings for the questions in each dataset
  • Finds a number of nearest neighbours in the examples_dataset for each entry in the input_dataset
  • Runs the guidance program on each question in the input_dataset, providing appropriate examples from the examples_dataset
  • Cleans up its intermediate keys from the result