Pragma to identify non-singleton classes

Motivation

We've seen a few places recently where it would be useful to mark classes as non-singletons, e.g. reflection could define class {-# NON_SINGLETON #-} Given a.

This would mean:

• The specialiser would never specialise for the class (but might abstract over free dictionary variables of that type instead, see #23282 (comment 500933)).
• Shortcut solving (-fsolve-constant-dicts) would be disabled for the class.
• Hence withDict should work reliably for such classes (see #21575 (comment 520219)).
• Using incoherent instances for the class would be less risky, because once type inference had determined an instance it would not change during optimisation (see #23172 (comment 500109)).
• This would provide a more robust way to prevent specialisation of incoherent instances than attaching nospec with -fno-specialise-incoherents as introduced in !10748 (merged) (see #23429 (closed), #23560).

The IP class backing implicit parameters effectively already has this behaviour, as I understand it.

We should more clearly document that in the absence of the pragma, the optimiser will assume classes are singletons, and hence incoherent instances / withDict are potentially unsafe.

Open questions

• Would this suffice for reflection? As far as I can see it would merely need to mark the Reifies and Given classes as non-singletons, but I'm not sure if this might end up having a performance impact. Perhaps @RyanGlScott could comment?
• Would this suffice for @cactus' use case?
• I'm unsure of the best concrete syntax for this, are pragmas or modifiers the right approach?
• Perhaps GHC should emit a warning or error if withDict was used for a normal non-singleton class?