[llvm] Do not use R12 for indirect tail calls with PACBTI (PR #82661)

[via llvm-commits]

ostannard wrote:

It looks like clang emits a module flag instead of a function attribute, so it applies to all functions in the module:
```
!4 = !{i32 8, !"sign-return-address", i32 1}
```

The attribute and flag are read by the `GetSignReturnAddress` function in `ARMMachineFunctionInfo.cpp`, with the function attribute taking priority over the module flag if both are present. This seems to work correctly for normal code-generation, so I'd expect it to work for these new instruction selection patterns too. Are you saying that this does not work if you test it end-to-end with clang, or when generating IR with clang?

For #81020, do you have a reproducer showing it not working correctly? I've tried a few simple examples and it seems to be working correctly, despite clang using module flags instead of function attributes:
```c
typedef void(func)(void);

void foo(func *fptr_arg) {
  register func *fptr_x16 asm("x16") = fptr_arg;
  asm("" : "+r" (fptr_x16) : : "lr");
  fptr_x16();
}
```

```
$ /work/llvm/build/bin/clang --target=aarch64-none-elf -march=armv8.3-a -c indirect-tail.c -o - -S -O1 -mbranch-protection=pac-ret+pc | sed -n '/foo:/,/func_end/p'
foo:                                    // @foo
// %bb.0:                               // %entry
        hint    #39
.Ltmp0:
        paciasp
        str     x30, [sp, #-16]!                // 8-byte Folded Spill
        mov     x16, x0
        //APP
        //NO_APP
        mov     x0, x16
        ldr     x30, [sp], #16                  // 8-byte Folded Reload
        adr     x16, .Ltmp0
        hint    #39
        autiasp
        br      x0
.Lfunc_end0:
```

https://github.com/llvm/llvm-project/pull/82661