Generate obj instead asm when lower llir to bin

[python3kgae]

Currently, _llir_to_bin generates ASM instead of OBJ

This approach works fine for Linux. However, for Windows using MSVC, compiling an ASM file and a CPP file together is not supported. You have to compile ASM separately and then link the resulting OBJ when compiling the CPP file.

I'm wondering if it would be acceptable to let _llir_to_bin generate OBJ directly. This would be similar to make_hsaco for AMD and make_cubin for CUDA, where the result is binary instead of text as well.

[nhat-nguyen]

@zhaoshiz does your team at Qualcomm depend on at all? would be great to hear your thoughts.

[zhaoshiz]

No we don't rely on that. We use our in-house compilation flow after triton-shared-opt.

[mitekoth]

It's as @zhaoshiz mentioned. We only use the triton-shared-opt directly for our use case as well.

[nhat-nguyen]

Thanks folks for confirming. We went ahead with the changes :)