Make `@testitem`s more isolated by isolating new methods/types

We recently saw a case (at RAI) where a testitem that reflected over `subtypes(AbstractFoo)` failed if it happened to run after (and on the same worker as) a seemingly unrelated testitem that defined a new `struct Foo <: AbstractFoo end`

Ideally it wouldn't be possible for testitems to interact in this way.

See also #184 