Restore FlatMap.map2Eval laziness

[satorg]

Resolves #4858.

[johnynek]

I guess the motivation was that inside the flatMap the .value might not blow the stack and if the F[_] is lazy it would possibly be safe.

The change makes us always evaluate the Eval[F[B]], but it is safe.

The motivation of the map2Eval is for Monads like Try/Either that a failure on the left avoids evaluating the right. This change will evaluate the Eval in those cases and no longer short circuit.

Maybe to mitigate this we want to make sure any shortcircuiting Monads in cats override map2Eval so we don't regress on cases like Try, Future, Either. Maybe others.

A middle ground could be to override StackSafeMonad to use the .value on the idea that the flatMap is always lazy in a StackSafeMonad. That may work or may have other issues that I don't see.

[satorg]

Maybe to mitigate this we want to make sure any shortcircuiting Monads in cats override map2Eval so we don't regress on cases like Try, Future, Either. Maybe others.

Yes, good call. I think that would a correct direction to go. We don't know and can't guess whether current FlatMap instance is for a short-circuiting Monad or not. The goal for generic methods is to be as safe as possible (I believe so). The concrete implementation though does know whether their Monad can short-circuit or not. So, it is more likely their responsibility to optimize for short-circuiting cases.