Allow explicit UDF arg & return types specifications

[sfc-gh-halu]

Two use cases:

• MLflow packaged models do not enforce the specification of column specs. Currently, we bail in this case. Would be good if customer could manually specify.
• MLflow auto schema capture could result in numpy type of object which could be because of python list, dict or even str. There's no way we could generate proper UDF signature based on object type. (reported by @jonwiggins)