[GPU] exchange LLVM optimization pass apply to GPU instead of host

[BI71317]

fixes #829

Changes

• updates the NVPTX optimization pipeline so that gpuopt1 and gpuopt2 run on the prepared GPU module instead of the host module.

Before:

runLLVMOptimizationPasses(module, /*plugins=*/nullptr, &gpuOptions);

After:

runLLVMOptimizationPasses(GPUmodule.get(), /*plugins=*/nullptr, &gpuOptions);

MRE

$ /usr/local/cuda-12.3/bin/ptxas -arch=sm_50 -v 02_gpu_exp_on_expr_with_rel.ptx -o 02_gpu_exp_on_expr_with_rel_no_mo.cubin
ptxas info    : 0 bytes gmem
ptxas info    : Compiling entry function 'exp_kernel_naver_02_0_0_std_numpy_ndarray_ndarray_0_float32_1__std_numpy_ndarray_ndarray_0_float32_1__' for 'sm_50'
ptxas info    : Function properties for exp_kernel_naver_02_0_0_std_numpy_ndarray_ndarray_0_float32_1__std_numpy_ndarray_ndarray_0_float32_1__
    8 bytes stack frame, 0 bytes spill stores, 0 bytes spill loads
ptxas info    : Used 29 registers, 368 bytes cmem[0], 16 bytes cmem[2]

Now with release mode, codon pipeline generates valid ptx.

Result

$ ./02_gpu_exp_on_expr 
4 array of float32
╭─────────┬─────────┬─────────┬─────────╮
│ 0.3679  │ 1.      │ 0.1353  │ 2.7183  │
╰─────────┴─────────┴─────────┴─────────╯

[arshajii]

Thanks, LGTM