Copy "seen" before handing it to resolve_schema_references to prevent…

[jorenretel]

Hi Jonvet,

thanks for solving this problem! I ran into it as well and found your PR (tadata-org#156). When I used your proposed fix though I found that there is one specific case where the current version of the code makes a mistake. It happens when a model has two properties on the same level that are of the self-referencing type. When this happens the second (and following...) properties are treated differently than the first one although they are at the same level.

Adding a copy to the seen set before handing it to new passes of resolve_schema_references makes sure that this does not happen as the first instance is not updating the parent set.

An example where this would happen:

class Player(BaseModel):
    name: str
    number_of_games: int
    teacher: Optional[Player] = None  # self-referencing

class Game(BaseModel):
    playerA: Player
    playerB: Player

In this case playerB will be represented by the $ref while playerA is allowed one level of recursion.