Fix a bug where this SDK's internals fail to handle items restored using Pydantic etc.

[seratch]

This pull request resolves a known issue with serialized/deserialized items.

The most common one is TypeError: cannot pickle 'pydantic_core._pydantic_core.ValidatorIterator' object with the data loaded via Pydantic methods, but we've received similar reports with other objects too.

This pull request minimizes the use of copy.deepcopy() inside this SDK to eliminate the potential runtime errors.

[rm-openai]

do we really need such a complex change? Can't we just replace copy.deepcopy(foo) with list(foo) to make a copy of the container?

[seratch]

@rm-openai yeah, we can simplify the implementation. I think I can remove the memo and deepcopy for basic objects and primitives. anything else you want to remove?

[rm-openai]

I think we don't need anything except replace all copy.deepcopy(...) with list(...)

[seratch]

@rm-openai Thanks for clarifying. Since we'd been using copy.deepcopy for union type values (str | list) and potentially we may encounter this issue with other types such as set, tuple. So, I think having this simplified common utility should be helpful for us to more easily avoid making the same mistakes in the future.

That said, if you prefer a different implementation, you can come up with your own changes. I will be happy to be a reviewer instead.

[rm-openai]

Yeah I think this PR is simpler: #1490. For example, it doesn't recursively copy everything in the list, because the only thing we care about is making a new top level list so that mutations dont affect the original list.