Add managed memory

[csbnw]

Description

There are many ways to mix and match different types of CUDA memory. As preparation for this PR, an (performance) evaluation of different strategies was conducted, the findings are summarized below:

1. The use of cu::HostMemory and cu::DeviceMemory with explicit memory copies provides the best performance, but is also the most verbose.
2. Mapped memory (already supported in cudawrappers) allows addressing host memory on the GPU, which allows for simple code, but it performs rather poorly.
3. Managed memory (also known as unified memory), provides better performance and given the user the option to control data movement by prefetching to either host or device memory. When like option 1, performance is only slightly lower than.
   With the changes in this MR, support for option 3 is added.

To be specific, the cu::DeviceMemory constructor can now also be used to allocate managed memory by passing CU_MEMORYTYPE_UNIFIED as CUmemorytype argument and optionally also some flags. This change is transparent to pre-existing code by having CU_MEMORYTYPE_DEVICE as the default CUmemorytype and default flags = 0.
Additionally, cu::Stream::memPrefetchAsync is added to expose the cuMemPrefetchAsync function.

The new functionality is tested in new sections of the test_vector_add test.

Related issues:

• Add support for registered and managed memory #173

Instructions to review the pull request

• Check that CHANGELOG.md has been updated if necessary

[john-romein]

A few minor things:

• in the constructor, the assignment to manager can be done outside (after) the if-then block
• there should be a checkCudaCall() around the call to cuMemPrefetchAsync()

[csbnw]

@john-romein,
I applied your suggestions 👍

[matmanc]

Looks really nice :)