Note [Determining whether newtype-deriving is appropriate]

When we see
  newtype NT = MkNT Foo
    deriving C
we have to decide how to perform the deriving. Do we do newtype deriving,
or do we do normal deriving? In general, we prefer to do newtype deriving
wherever possible. So, we try newtype deriving unless there's a glaring
reason not to.

"Glaring reasons not to" include trying to derive a class for which a
coercion-based instance doesn't make sense. These classes are listed in
the definition of non_coercible_class. They include Show (since it must
show the name of the datatype) and Traversable (since a coercion-based
Traversable instance is ill-roled).

However, non_coercible_class is ignored if the user explicitly requests
to derive an instance with GeneralizedNewtypeDeriving using the newtype
deriving strategy. In such a scenario, GHC will unquestioningly try to
derive the instance via coercions (even if the final generated code is
ill-roled!). See Note [Deriving strategies].

Note that newtype deriving might fail, even after we commit to it. This
is because the derived instance uses `coerce`, which must satisfy its
`Coercible` constraint. This is different than other deriving scenarios,
where we're sure that the resulting instance will type-check.