Use correct CFA register on ARM64

[NickBarnes]

I noticed that testsuite/tests/unwind/driver.ml was failing on my Mac. @tmcgilchrist identified and fixed the problem (wrong CFA register), which was related to my use of an old version of XCode, not exercised by the CI. This solution is his but with a tweak from me for the check-linker-version.sh script.

[NickBarnes]

Updated to remove a CI workflow change which was breaking the MacOS CI. https://github.com/ocaml/ocaml/actions/runs/11720868392/job/32647116985

[tmcgilchrist]

This works well for Xcode versions 14 through to 16 on both ARM64 and x86_64.
Originally this bug (which I wrote) was missed because the test suite was incorrectly skipping this test.

[ghost]

FWIW, asmcomp/riscv/emit.mlp has a similar bug, where the cfi_def_cfa_register stanza on line 342 refers to register 21 which is the correct number for s0 in int_reg_name but is the wrong DWARF number for it, which is 8.

Edit: per @tmcgilchrist's proposal, submitted as separate PR: #13607.

[dra27]

Nit in the bash script - I haven't actually reviewed

[gasche]

I morally approve of this change (even though I don't know anything about 29 and 19 as magic numbers for ARM64). I would suggest that you credit one of yourselves (Tim or Nick) as reviewers, and then we merge.

[tmcgilchrist]

@gasche I introduced the error in #13500. x29 is the frame pointer register on ARM64. Copying my comment from that PR explains why x19 matters.

22fee2d Stores sp in temporary register x19 during Iextcall in the same way as amd64 backend. Previously this was implemented in #13079 as push/pop on the stack, however this interferes with frame pointers.

My analysis is this test was failing because the logic for parsing the ld -v response was incorrect and it wasn't noticed when I submitted the Frame Pointer work.