Fix miscompilation / liveness errors for string operations [superceded]

[mshinwell]

This fixes the random CI failures seen on arm64.

It turns out that the Cmm machtypes given to "pointer + offset" calculations used for all of the different varieties of string reads and writes are wrong. Some of them have been wrong for a very long time (30 years). They are currently given type Int, but in the case where the pointers may be into the OCaml heap (e.g. for string and bytes rather than Bigstring.t) they must have type Addr. Otherwise, CSE may apply, causing pointers into the middle of values to be live across a GC. For example in warnings.ml:

  and loop_letter_num tokens modifier i =
    if i >= String.length s then error () else
    match s.[i] with  (* first string access *)
    | '0' .. '9' ->
        let i, n1, n2 = get_range i in
        loop (Num(n1,n2,modifier)::tokens) i
    | 'A' .. 'Z' | 'a' .. 'z' ->
       loop (Letter(s.[i],Some modifier)::tokens) (i+1)  (* repeated string access after potential GC *)
    | _ -> error ()

This patch affects the system compiler as well as the flambda-backend compiler, hence the use of a patch file to fix the former. I will deal with getting this fixed upstream. The system compiler patch isn't currently as good as the flambda-backend patch since it doesn't special-case out-of-heap accesses.

There are a couple of other changes currently in here too, which can be discussed and maybe split:

• fixes to stop the system compiler applying Comballoc across Ccheckbound, which can raise an exception. I need to discuss this with @stedolan though, maybe this is only required if the allocation is on the major heap? (This problem does not arise with the flambda-backend compiler since bounds checks are expanded before Flambda 2.)
• fixes to silence the debug runtime with the system compiler, which I think was causing excessive CI logs in some cases.

This bug was very troublesome to track down and @xclerc must be thanked for his help with this, especially since we thought there was a second bug, which was actually caused by a mistake in my hack for tracking down the first bug! Some of the CI changes in this PR were written by @xclerc .

Once the dust has settled I will re-enable the arm64 builds as required.

[mshinwell]

Something is wrong with the CI here, re-opening.

[mshinwell]

See #3960