Annotate nullable values as such in the OpenAPI schema

[erbridge]

Prior to this change, we allowed the return of None values from models, serializing them as null, but the schema didn't specify them as nullable, which would cause model output to fail validation.

This only supports Pydantic V2, because, from what I can tell, Pydantic V1 doesn't give us any clues that a field allows None values.

Note: this is probably a breaking change.

[aron]

I think this is good to merge (bar the input_seed.py file).

Regarding whether this is a breaking change. From what I understand, optional types will now include the nullable field on the OpenAPI schema for both inputs and outputs. In practice this allows a client to pass field: null in addition to omitting it from the request body, it also correctly types the output to indicate that output can be null, this has always been the case when a prediction is in a non-terminal state.

The only two failure modes I can think of are:

1. Where someone is using the generated OpenAPI schema for something else. The new presence of the nullable field could change the behavior of whatever system is using the schema.
2. When the output schema is used to determine the type of the output. However this has always been a nullable field on the PredictionResponse so I would expect most consumers to handle this already despite the incorrect schema.

I think based on this we can class it as a bug fix without a major version bump.

[8W9aG]

I think this is good to merge (bar the input_seed.py file).

Regarding whether this is a breaking change. From what I understand, optional types will now include the nullable field on the OpenAPI schema for both inputs and outputs. In practice this allows a client to pass field: null in addition to omitting it from the request body, it also correctly types the output to indicate that output can be null, this has always been the case when a prediction is in a non-terminal state.

The only two failure modes I can think of are:

1. Where someone is using the generated OpenAPI schema for something else. The new presence of the nullable field could change the behavior of whatever system is using the schema.
2. When the output schema is used to determine the type of the output. However this has always been a nullable field on the PredictionResponse so I would expect most consumers to handle this already despite the incorrect schema.

I think based on this we can class it as a bug fix without a major version bump.

Luckily we have another breaking change going in right now, the next bump will be to 0.15.0 so we can make this part of that bump.

[8W9aG]

Going to merge this since we need to do a 0.15.0 release and I believe all the comments have been taken care of