[Doc] Document the **why** for the different parts of AMLTK

[eddiebergman]

Right now, the documentation really focuses on that how, i.e. how you can do A, B, C but not why you might want to use AMLTK.

The why you might want to use these components is completely missing.

Thanks to @vladislavalerievich for making this point clear!

---

For examples:

• Why do we need these Component, Choice etc...
  • Define your pipeline and search space in one place
  • Easily get an overview of the pipeline due to its declartive nature
• Why do we want to write things using the Scheduler
  • Local experiments on processes is easy, however the same code that works locally can't scale up to a cluster without making changes to the code or having lots of branches. The Scheduler allows the same code to run locally and on cluster scale without changes.
• Why do we use Trial and Optimizer?
  • There are usually a lot of things that need to be stored and the changes to your evlauation function will change dynamically during development. Making these changes should be quick and easy. e.g. logging new things in Weights and Biases is easy and it stores lots of additional information which might be useful later on, before being thought about. The Trial also helps with bookkeeping, i.e. a unique directory per trial and also making it easy to store artifacts per trial.
  • The Optimizer interface and the provided interfaces basically make access to different Optimizers uniform. Otherwise, you'd have to deal with a lot of changes going from one Optimizer to a new one.