Make sure that our pickling process is robust against future changes

[yifanwu]

Right now we rely a lot of pickled columns via SQLAlchemy (via type decorators).

Changes include:

• Python minor version change
• SQLAlchemy versions
• Database versions
• Database types
• What machine things are ran on

It's probably important before we start supporting real customers to make sure that the pickling process is robust and backward compatible.

I think strings are more robust against these software changes but might be slower? Something to consider.

[dorx]

For some context on pickler choice: apache/airflow#7870

We should use cloudpickle if it's something that airflow et al. is using.

We can pass in our pickler of choice via:

We would then instantiate the column with something like:

Column(PickleType(pickler=cloudpickle.pickle), nullable=True)