[REVIEW] Add CUDA Managed Memory (Unified Memory) Allocation Mode

[harrism]

This PR adds an option to rmmInitialize to use cudaMallocManaged (+ cudaMemPrefetchAsync) as the base allocator for RMM rather than cudaMalloc.

rmmAllocationMode_t mode = CudaManagedMemory | PoolAllocation;
rmmOptions_t options {mode, 0, false};
rmmInitialize(&options);

The code above will initialize RMM with a managed memory pool. Removing PoolAllocation will cause it to allocate without a pool by calling directly to cudaMallocManaged().

Tests are also expanded to run all tests with all allocation modes (gtests and pytests).

Note this depends on new functionality I added to cnmem with this PR (already merged): NVIDIA/cnmem#7

This PR is an improved version of a (never merged) cuDF PR rapidsai/cudf#422.

Support for multiple simultaneous pools with different modes is future work.

[jrhemstad]

Only major change is that rmm::alloc needs to be updated to use cudaMallocManaged if CudaManagedMemory is enabled, but PoolAllocation is not.

[jrhemstad]

Good to go after these changes.