[Diagnostics] Add a fix-it to insert 'some' when associated type inference failed because existential types can't conform to protocols.

[hborla]

Protocol conformances for existential types are not supported today. If a programmers tries to fulfill an associated type with conformance requirements using an existential type, they'll get an unactionable error message:

protocol P1 {}
protocol P2 {
  associatedtype A: P1 // note: unable to infer associated type 'A' for protocol 'P2'
  var a: A { get }
}

struct S1: P1 {}
struct S2: P2 { // error: Type 'S2' does not conform to protocol 'P2'
  var a: P1 { S1() } // note: Candidate can not infer 'A' = 'P1' because 'P1' is not a nominal type and so can't conform to 'P1'
}

When the associated type was inferred from an existential result type, the compiler can suggest that the programmer use an opaque type instead by inserting some, which is usually a better solution for hiding a concrete return type behind a protocol:

note: cannot infer 'A' = 'P1' because 'P1' as a type cannot conform to protocols; did you mean to use an opaque result type?
  var a: P1 { S1() }
         ^~
         some 

[slavapestov]

This is an excellent idea.

[slavapestov]

Nitpick: what about SubscriptDecl?

[hborla]

@swift-ci please smoke test