Unify unsized types in type inference

[nrc]

In order not have a tonne of coherence errors, we currently do not unify unsized types with type variables in type inference. For example, [T] with U. This mimics our old behaviour where unsized types were not types.

This prevents us having an impl for &T which will use an impl for [T], i.e., we need a separate impl for &[T]. This is one of the primary benefits of DST. It also causes problems when the compiler needs to infer a DST for a type parameter, e.g., in function calls. See #17122 for one example.

Blocks #16918, #12938

[nrc]

Nominating for p-backcompat-lang. This is the last backwards incompatible part of DST, so we can unblock #12938 from 1.0. This change is backwards incompatible since it will introduce coherence errors for impls for both sized and unsized types.

[nrc]

Plan A:

• modified variation on trait-reform nested obligation rules (consider the bounds of T in &T so that &T where T is Sized vs not Sized are different types)
• make snapshot
• update impls
• patch up trait-reform nested obligation rules

Plan B

• hack up type inference to track whether the parameter was Sized?
• prevent unification if not
• as above

[pnkfelix]

Assigning P-backcompat-lang, 1.0 milestone.