Compiler: narrow `interp` in `typeinf` loop (#63634)
interp = callee.interp` read the abstractly typed field, so the work
loop dispatched on `AbstractInterpreter` while the frame was concrete.
With more than one interpreter loaded (REPL's `REPLInterpreter`),
inference then compiled every interpreter/state combination, e.g.
`abstract_call_known(::REPLInterpreter, ..., ::InferenceState{NativeInterpreter})`.
Narrow it with `::I`. To not box the now concrete interpreter again, pass
the field to the dynamic `task(...)` call in `doworkloop` and avoid
`applicable`/`===` against the abstract field in the `Future` asserts.
Keep `typeinf_local` behind a `compilerbarrier`, called with the boxed
field. A static call folds it and everything it reaches into one
inference cycle with `typeinf`, and cycle members are neither inlined
into each other nor called through their specsig, so e.g. `abstract_call`
boxed the interpreter on every call.
Add REPL completion warmups for splatting and `invoke`, which the mixed
specializations used to cover by accident.
