If someone wanted to use Hamilton to model a modeling dataflow,
they would struggle. Need a new decorator to handle extracting
outputs from functions that return multiple things and aren't
a dataframe.

The API to use it looks like this:

```python

@function_modifiers.extract_fields(
    {'X_train': np.ndarray, 'X_test': np.ndarray, 'y_train': np.ndarray, 'y_test': np.ndarray})
def train_test_split_func( ... ,  ... ) -> dict
```

I decided to go with a straight dict of `field_name` to `field_type` because that seemed
the simplest thing to define. Note, we use the documentation for the original function,
rather than enabling individual doc strings for the types. I think this suffices for now.

To support TypedDict, I didn't want to have to import typing_extensions to handle
it. Also you can't inline define a TypedDict class, so it would be more verbose which
is less that ideal.

We can always add TypedDict support later. Also punted on `Tuple` support -- that
might be another decorator...

This example is interesting because it shows how one might
build a "bank" of hamilton functions that do some generic
modeling -- while keeping it generic so that adding new contexts/
running it with new models, results in a small amount of work.

The key things to get this to work are:

 - different python modules to load data. They have to output what's required to link
   with the my_train_evaluate_logic functions.
 - config & @config.when to add the correct model function to the dataflow.

So if you want to switch between model types: easy -- change config.
So if you want to switch fitting models on different data: easy -- change the data loading module.

Helps prove things work as intended!

I wonder if this is flakey somehow? Anyway adding this to see if circleci complains
or not. Seems like there could be a version mismatch somewhere that causes this,
i.e. my local env, versus what circleci installs, etc.