[llvm] [AMDGPU][GlobalISel] Enable kernel argument preloading (PR #134655)

[Tim Gymnich via llvm-commits]

tgymnich wrote:

@arsenm do we want to skip kern arg preloading if the argument is not used? This kind of makes the tests in their current form a bit pointless.

In that case we might want to test with this func instead:
```llvm
define amdgpu_kernel void @amdhsa_kernarg_preload_4_implicit_6(i128 inreg %x, ptr addrspace(1) %p) {
    %res = add i128 %x, %x
    store i128 %res, ptr addrspace(1) %p
    ret void
}
```

https://github.com/llvm/llvm-project/pull/134655