Fixed LibDevice compilation on compute_100 and later.

[MoFtZ]

A member of the community reported issues using LibDevice on a newer RTX 5090 device.

After some investigation, the NVVM compiler returned NVVM_ERROR_COMPILATION on the NVVM IR:

!nvvmir.version = !{{!0}}
!0 = !{{i32 2, i32 0}}

target triple = \"nvptx64-unknown-cuda\"
target datalayout = \"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64\"

declare float @__nv_cosf(float %x)

define float @__ilgpu__nv_cosf(float %x)
{
entry:
    %call = call float @__nv_cosf(float %x)
    ret float %call
}

This could be reproduced by setting the compiler argument -arch=compute_100, or newer.
https://developer.nvidia.com/cuda-gpus#:~:text=Table_title:%20CUDA%20GPU%20Compute%20Capability%20Table_content:%20header:,NVIDIA%20A100%20NVIDIA%20A30%20%7C%20GeForce/RTX:%20%7C

After investigating further, the solution is to place the target directives first.

I have also updated the target datalayout directive with the i128 registers, as per the NVVM IR specification:
https://docs.nvidia.com/cuda/nvvm-ir-spec/index.html#data-layout