Polls in loops can create control flow invisible at the Flambda level

[lthls]

I'm creating an issue corresponding to ocaml#10520 in the upstream compiler.
Flambda would be allowed to do the same optimisations that Deadcode does, removing the initialisation step for mutable variables that are always assigned to before being read. If a poll is later inserted in the loop, this could introduce a path where the initialisation step becomes necessary again.

This is not a problem now, but will become one when we rebase on 4.13.

[mshinwell]

I talked to @stedolan about this. One reasonable solution might be to have a "maybe poll" operation in Flambda 2. This would be inserted at a superset of points where Polling might put the Ipoll instruction. Since Polling only inserts polls in a very small number of different kinds of places, this should be feasible.