Select better auto-property backing field type to support devirtualization

[AustinWise]

Fixes #30797

[jaredpar]

Thanks for the contribution here. Couple of things:

1. We still need to clarify whether or not we want to take this change with the language design team. I'm optimistic this will be supported but I need to ensure that's the case.
2. Tests: with this change I would expect to see a number of tests here, both positive and negative. There don't seem to be any tests so far.

[AustinWise]

I got this working in the editor experience and the compiler. There is still some work to do before merging, but I'd like some feedback to make sure this approach is not totally wrong.

Things I think I need to do before merging:

• Add unit tests
• Perhaps refactor the binder code so that BindVariableOrAutoPropInitializerValue and this new code share the same logic for determining field conversions.

[AustinWise]

Rebased and updated with some unit tests. These mostly focus on the positive cases (implicit reference conversions from the initializer expression to the property type).

[jaredpar]

In the process of checking with LDM about this particular change we ended up discovering some problems with the proposal. I added a lot of detail to the original issue + CC'd you.

#30797

[AustinWise]

As mentioned in #30797, the fact that constructors can reassign readonly auto properties make this approach intractable.

Interesting, there is no test currently that excises this feature in a way that was broken by my changes. Perhaps it would be worth adding a little unit test to exercise this case:

class A { }
class B1 : A { }
class B2 : A { }
class C
{
    public C()
    {
        P = new B2();
    }
    A P { get; } = new B1();
}

[jaredpar]

I agree. I was surprised as well. I got lulled into a false sense of completeness when your PR passed the first round of testing.

Again thanks for trying to get this optimization in! Dissapointing that it's harder than we thought but also good to have a firm resolution on it.