Otherwise, the generated bytecode is unnecessarily suboptimal in some
(arguably weird) cases.
In the JVM backend, this was an accidental regression in #3260, as I had
not noticed that effectively inline-only functions were handled by a
separate branch in FunctionCodegen. In JVM_IR, I'm pretty sure the
redundant markers have always been there as `isSuspensionPoint` in
ExpressionCodegen never checked for effectively-inline-only-ness.