add support for CUDA allocation flags

[bratpiorka]

Add support for CUDA allocation flags.

This PR adds one new API call for CUDA:

• umfCUDAMemoryProviderParamsSetAllocFlags()

Checklist

• Code compiles without errors locally
• All tests pass locally
• CI workflows execute properly

[vinser52]

According to the CUDA Driver API spec, these flags are defined as follows:

#define CU_MEMHOSTALLOC_PORTABLE 0x01
#define CU_MEMHOSTALLOC_DEVICEMAP 0x02
#define CU_MEMHOSTALLOC_WRITECOMBINED 0x04

and

enum CUmemAttach_flags {
    CU_MEM_ATTACH_GLOBAL = 0x1,
    CU_MEM_ATTACH_HOST = 0x2,
    CU_MEM_ATTACH_SINGLE = 0x4
};

We cannot catch the situation when cu_params->memory_type = UMF_MEMORY_TYPE_HOST and cu_params->alloc_flags = CU_MEM_ATTACH_GLOBAL because CU_MEM_ATTACH_GLOBAL == CU_MEMHOSTALLOC_PORTABLE

I think, we can do nothing, it is a user's responsibility to provide a correct combination of memory type and allocation flags.

So I leave this comment just to note.

[ldrumm]

I agree. We shouldn't try and catch invalid combinations of flags since CUDA will always be more up to date than us and will do the error checking for us. As long as errors from the CUDA runtime are correctly propagated I think we should just pass through

[bratpiorka]

you are right, I have removed this check and added a comment

[igchor]

Hm, I think it would be better not to modify the params. In theory, following scenario is valid, right?

auto params = umfCUDAMemoryProviderParamsCreate();
params.setMemType(UMF_MEMORY_TYPE_SHARED);

provider1 = umfMemoryProviderCreate(..., params);

params.setMemType(UMF_MEMORY_TYPE_DEVICE);
provider2 = umfMemoryProviderCreate(..., params);

In this case, the alloc_flags passed for the second provider will be different than what user would expect.

[vinser52]

Good catch. We should do this for the cu_provider->alloc_flags.
BTW can we test such scenario in our tests?