Fix typevar variance for agent deps

[dmontagu]

Right now, the variance of RunContext is the reverse of what it should be, because we are reusing a typevar rather than creating a type-specific one. Because of this, you can't define reusable tools that use deps but aren't exact matches on the agent deps type. This change fixes the typing (it is actually incorrect, this is a typing bugfix not just a change), and in particular makes it work to define tools as accepting supertypes of the agent's actual deps class, as you can see in the new typed_deps.py file I added as a typing test.

(I recall pointing out the potential for reusing typevars to cause issues when we first did it; it won't surprise me if it remains an issue. In general I would strongly prefer to use infer_variance if it wasn't for the fact that it makes it impossible to have dataclasses in 3.13 that are covariant due to the __replace__ method.)

This addresses the issue described in this thread: https://pydanticlogfire.slack.com/archives/C083V7PMHHA/p1762185662720909

[github-actions]

Docs Preview

commit:	7ab0a5b
Preview URL:	https://9b2a62a3-pydantic-ai-previews.pydantic.workers.dev