Capture the ambient discussion and thinking about bi-directional type information flow in a principle

[chandlerc]

Several design discussions in Carbon have talked about "what direction does type information flow" in the context of type inference. Broadly, these discussions have seemed to advocate away from bi-directional type information flow.

More specifically, they have advocated away from "inwards" flow such as the context using the result of a call expression (as opposed to the arguments) influencing overload selection: https://github.com/carbon-language/carbon-lang/blob/trunk/proposals/p2875.md#overloaded-call-operator

Also more specifically, they have moved us away from a Hindley–Milner type system design for expression types, etc.

We should capture the underlying idea that informs these discussions as best we can and write it into a principle document that can be referenced. A nice property of such a principle is it can be relatively brief and doesn't need to be precise or exacting, but can still help indicate the directionality of Carbon's design in this space.

[chandlerc]

(marked as long-term because this isn't urgent at all, mostly filing the issue to not lose track of it)