Refactor to make Pipeline / make elm better for interactive use

[PeterDSteinberg]

In working on the Loikith example problem, I found some of the configuration of training in ensemble unnecessarily wordy and confusing. This PR aims to fix that by:

• Better separating the config parsing logic from the machine learning logic
• Making a Pipeline object similar to that of sklearn.pipeline. See:
  • elm.pipeline.pipeline: The Pipeline class
  • elm.pipeline.steps: All the classes representing steps that can be used in a Pipeline

How this Pipeline differs from that of sklearn.pipeline:

• elm.pipeline.Pipeline allows passing / transform X, y and sample_weight data in each step, where X is an ElmStore or xarray.Dataset.
• Where sklearn uses a FunctionTransformer as class for a user-given function that transforms X, we have elm.pipeline.steps.ModifySample. See its usage in elm.pipeline.tests.test_pipeline

[PeterDSteinberg]

Fixes #88 Fixes #66 Fixes #59

[PeterDSteinberg]

Fixes #68