Support generic upper bounds

Feature

Allow generic upper bounds such as:

class ClassA[T, S: Sequence[T]]: ...

Pitch

PEP 695 has an example that currently doesn't work, and alludes to a future extension to the typesystem that eliminates the limitation:

# The following generates no compiler error, but a type checker
# should generate an error because an upper bound type must be concrete,
# and ``Sequence[S]`` is generic. Future extensions to the type system may
# eliminate this limitation.
class ClassA[S, T: Sequence[S]]: ...

I ran into a usecase for this that I have no way of expressing with the current generics limitations, and generic upper bounds would solve it. I will provide a simpler version of it here so it's easier to discuss.

from typing import Callable, Iterable, Sequence

type CostFunction[T] = Callable[[T], float]

# Sequence[T] gives an error, because T is not concrete
def get_lowest_cost_sequence[T, S: Sequence[T]](
    seqs: Iterable[S], item_cost: CostFunction[T]
) -> S:
    def seq_cost(seq: S) -> float:
        return sum(item_cost(t) for t in seq)

    return sorted(seqs, key=seq_cost)[0]

In the example, we are working with sequences, and a cost function for each item in the sequences. The crux is that the function must return the same type as the original sequences. In other words this is an incorrect type signature:

# Return type Sequence[T] is looser, because it does not mean whatever is
# returned will be _the same type_ as what is passed in. Any users of this
# function would have to cast the result for it to expose the same functionality
# as what was in `seqs`.
def get_lowest_cost_sequence[T](
    seqs: Iterable[Sequence[T]], item_cost: CostFunction[T]
) -> Sequence[T]:
    ...

Aucun fichier d’implémentation, test ou point d’entrée n’est indiqué. Commencez par la section « type-parameter-scopes » de PEP 695 et les exemples de cet Issue, puis suivez la gestion par mypy des bornes supérieures génériques ; le travail est terminé lorsque les bornes génériques valides sont acceptées tandis que les bornes non concrètes invalides restent rejetées, avec une couverture de régression.