Support different allocation modes in device allocator

[upsj]

CUDA and HIP provide functions for allocating unified memory (cudaMallocManaged) and pinned host memory (cudaMallocHost), which transparently allow access to both host and device. They might be interesting in application integrations (hybrid CPU/GPU workloads) and MPI (pinned host buffers if not GPU-aware). A first suggestion for the design would be providing _host and _unified (or _shared) overloads to (raw_)alloc and (raw_)free, as well as adding an enum class allocation_type to ExecutorDeleter. We could use another overload to gko::array taking this enum as a final parameter, to enable a straightforward integration into the rest of Ginkgo.

[MarcelKoch]

Currently, our cuda and hip executors can already allocate unified memory through the allocation_mode enum. By creating executors with different enums, it is also possible to switch between the different modes.

So, besides adding pinned host memory and DPCPP support, how would ginkgo, or a user of the library, benefit from these additions? Also, should these additional overloads replace the allocation_mode enum? I don't think having both makes sense.

[upsj]

should be fixed by #1315