Cold continuation handlers in flambda2

[Ekdohibs]

Cold continuation handlers will not be inlined, neither by simplify, nor by to_cmm. For now, they are generated only in from_lambda for the exception handlers of checked operations, but the plan is to add then in lambda as well and propagate the information. The backend parts seem more complex however: while it seems to be easy enough to propagate the information into Cmm code, it's unclear how to represent it inside Mach and Linear if we want to be able to put cold continuation handlers at the end of the function.

[Ekdohibs]

This now includes propagation of is_cold to Cmm and Mach, and is ready for review; backend support (at least Cfg) to ensure cold continuation handlers are compiled in a different section or at the end of the function will come in a separate PR.
Some things I'd like to be given some thought about this PR:

• Cmm and Mach handlers are becoming large tuples, and this PR needs to change every single place they are used causing a large diff. Should we change them to a record instead?
• Can we try to use the coldness information is spill/reload to say that the cost of reloading a register on cold paths is reduced?

[gretay-js]

• Cmm and Mach handlers are becoming large tuples, and this PR needs to change every single place they are used causing a large diff. Should we change them to a record instead?

I also think it should be a record, at least for the handlers. I think the change to a record will make the diff in this PR bigger though, so maybe a separate PR would actually be easier to review.

• Can we try to use the coldness information is spill/reload to say that the cost of reloading a register on cold paths is reduced?

yes, I think the cfg-based register allocators are well suited for using this info.
What would be particularly useful for spill/reload placement is to propagate to the backend which calls are cold, based on user-provided [@inline never][@specialise never] annotation or even better provide compiler suport for @cold attribute and propagate that.

[gretay-js]

only for files owned by me.

[Gbury]

I agree that we should introduce some records for the Cmm.Ccatch case because it's becoming way too big of a tuple/list nested type, and that it should be done in a separate PR.

[Gbury]

Good to merge, ^^