[BUG] Do not hardcode SM version

[yunwei37]

Currently we hardcode sm_60 in the cuda attach part. we should not do that

[Officeyutong]

It's ok but when don't have way to compile PTX in a higher version since we use LLVM

[yunwei37]

I remember we discussed some apis possible, can you put them here?

And, we can compile the trampoline code with llvm, but we should not compile the disassembled PTX code (and the eBPF code) with a fixed sm version.