Follow up issues for ml_tools

[lgray]

List of to-do's for #798 for after it is merged:

• use empty_if_typetracer / length_one_if_typetracer to remove need user to hint necessary columns
• use weakref/dask.delayed references to have fast tracing while robustly shipping large networks to processing cluster
• update documentation once empty_if_typetracer / length_one_typetracer solution available
• re-enable triton testing when dask_awkward can handle nested dask collections in kwargs (fix: adjustments to callable wrap to deal with typetracers in nested python structures #843)

@yimuchen

[nsmith-]

For the second bullet, let's try to avoid weakref whenever possible.

[lgray]

It's necessary if you want to do column_optimization locally and then also have efficient execution remotely.

We'll figure it out - I don't like the pattern either.

[lgray]

We now have a much better way to include ml models in the task graph that seems to not incur that much serialization.