Clean up parameter handling in forecasting pipeline

[wgifford]

• Cleans up defaults in init, using a similar approach to the audio pipelines in HF.
• Adds an option to control inverse scaling of outputs (driven by the time series preprocessor)
• Fixes some type hints

To do (possibly in a subsequent PR):
Better integration of the time series preprocessor in the pipeline.
User shouldn't have to pass all the parameters for columns multiple times, etc.
Use the preprocessor in the pipeline to actually preprocess and scale data!