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

[bratpiorka]

@ldrumm @igchor @vinser52 all tests pass, ready to re-review

[ldrumm]

Apologies for the delay. I should be able to properly review this later today.

[ldrumm]

I'm happy with this implementation of this feature as it pertains to CUDA, but do we need to better generalize for other accelerator APIs that also takes flags for allocations? CUDA is one, but HIP supports a different set of flags, and Intel OpenCL extensions also extend allocation to take hints.

As I discussed here I've already got work in progress that does similar things for all accelerator targets, so if UMF can provide a generic way to do this, then I'd prefer that

Please note, I'm still not to be considered competent with understanding how UMF works so feel free to tell me I'm misunderstanding

[ldrumm]

These are no longer doxygen style comments. Is that intended since this is implementation and non-public?

[bratpiorka]

Yes, this code was moved from the public include files to this location when we updated the UMF API to use setters instead of directly exposign the parameter structure. So this change is just a cleanup

[igchor]

I'm happy with this implementation of this feature as it pertains to CUDA, but do we need to better generalize for other accelerator APIs that also takes flags for allocations? CUDA is one, but HIP supports a different set of flags, and Intel OpenCL extensions also extend allocation to take hints.

As I discussed here I've already got work in progress that does similar things for all accelerator targets, so if UMF can provide a generic way to do this, then I'd prefer that

Please note, I'm still not to be considered competent with understanding how UMF works so feel free to tell me I'm misunderstanding

@ldrumm UMF doesn't allow flags/hints to be passed to pool allocation APIs. This is because those flags/hints can only be applied to the entire memory page. If we would, for example, allow umfPoolMalloc(1024, READ_MOSTLY_HINT), internally UMF would have to apply the hint to the entire memory page which would impact other, unrelated allocations.

Our current thinking is that the upper level (UR or SYCL) should decide what to do for every specific flag. For some flags (ones that are used often), it might make sense to create dedicated pools (for which entire memory is allocated with those flags). For others (or when the allocation size is big), it might be better to not do any pooling (just pass the request directly to the driver). UR spec gives as quite a lot of flexibility in how we can implement this in the adapters.