I compared Cfg_polling to Polling and they do the same thing, as far as I can tell, assuming correctness of Cfg_loop_infos.compute_back_edges, except on exceptional return from a function (easy fix, see comments).

• Can you please rebase this PR? The old diff confusingly includes #1927 that has been merged.
• Reminder about #2228. It becomes important for performance with runtime5 when we turn on poll insertion, as far as I remember from an upstream discussion (can't find the link, sorry). A small change around Cfg_to_linear.needs_starting_label is probably enough for backedges to implement it. If it's not obvious, we should at least have a comment there. We will of course do more benchmarking later, and if it turns out that this optimization is not significant, we can remove return_label from IPoll and simplify emit.mlp handling of IPoll.
• Please port comments such as this from polling.ml to cfg_polling.ml
• Does error handling need to be ported? I'm mostly worried about forgetting to copy report_error and register_error_of_exn over if we remove Polling . Maybe it's okay as is if there is a test that would fail. Alternatively, this PR can just move the shared code, including report_error, from Polling over to Cfg_polling.