[CIR][CUDA] FIx CUDA host compilation on kernel launch

[RiverDave]

This PR implements some missing blocks that allow us to effectively allow us to launch kernels from the host. All of the tests stated in this commit are now resolved.

I spent half a day figuring the following:

I tried experiementing performing host compilation(-fcuda-is-device) with target triple: nvptx64-nvidia-cuda but was getting a module verification error that, to keep it simple looked like: error: 'cir.call' op calling convention mismatch: expected ptx_kernel, but provided c.

I thought that was expected given that we're essentially using the device to compile on the host, which doesn't make a lot of sense. until I tried to replicate the same in OG and didn't really run into any problem in that regard. Are the calling conventions enforced in CIR much more strict as compared to OG? Or is that simply a bug from OG?

[bcardosolopes]

I thought that was expected given that we're essentially using the device to compile on the host, which doesn't make a lot of sense.

Interesting!

until I tried to replicate the same in OG and didn't really run into any problem in that regard. Are the calling conventions enforced in CIR much more strict as compared to OG? Or is that simply a bug from OG?

Seems like CIR is more strict, it's possible LLVM has some verifier issues or it could be a bug, I don't have (know) a good answer for you!

[bcardosolopes]

LGTM, thanks!