Type-unstable hot loops

Address type-unstable hot loops. Things worth trying, or order of invasiveness:

1. Introduce a function barrier. Replace

for d in devices # typeof(d) varies: loop body is type unstable.
    # code here won't specialize.
end

with

for d in devices
    _inner!(container, d, ...)
end

function _inner!(contrainer, d::D, ...) where {D}
    # compiled per concrete D: code here will specialize
    # (still a runtime dispatch, though)
end

If the loop body is sufficiently complex, and devices contains a couple different concrete types, the 2nd should win out. This may increase compile-time, though, so it's worth measuring.

2. Add a replacement for FlattenedIterator that unrolls across the different types. See IS #505. Requires interface work in IS (prototyped in that PR) plus rewriting the loops to use map and similar.