Simplify HostBuffer memory ownership

[nirandaperera]

HostBuffer tracked two redundant deallocation paths (mr_ + owned_storage_), where mr_ was stored unconditionally but never used when owned_storage_ was present — and as a non-owning ref it could silently outlive a PinnedMemoryResource.

Solution

• Replace both members with a single std::function<void(rmm::cuda_stream_view)> deallocate_fn_, invoked with the live stream at deallocation time to respect set_stream().
• In the RMM-allocated path, wrap mr in cuda::mr::any_resource before capturing it — deep-copying the resource and incrementing its refcount if it is a PinnedMemoryResource.
• Remove the vestigial mr parameter from from_owned_vector and from_rmm_device_buffer (it was always unused for deallocation).
• Remove the leaked OwnedStorageDeleter public type alias.

[madsbk]

I don’t think owned_mr actually owns the memory resource. AFAIK, cuda::mr::any_resource<P> only owns the rmm::host_async_resource_ref mr, which is itself a non-owning reference. So owned_mr ends up owning a copy of a non-owning reference, not the underlying memory resource?

[wence-]

This is a reasonable thought but incorrect. any_resource(resource_ref) magically converts the ref into an owning type.

[madsbk]

Thanks @wence-, good to know! @nirandaperera could you add a comment noting this?

[nirandaperera]

Yeah, any_resource takes in a copy of the underlying resource in the ref. Idea is to increment the refcount if that resource is shared.

[madsbk]

@nirandaperera, it would be good to add a test that constructs a HostBuffer from a transient memory resource, for example a PinnedMemoryResource that goes out of scope before the buffer is deallocated.

[nirandaperera]

@madsbk done

[nirandaperera]

/merge