[FEA] Set RMM async allocator as default

[rongou]

Is your feature request related to a problem? Please describe.
Currently the RMM arena allocator is the default, but in some circumstances it can still run into OOM errors due to memory fragmentation. The async allocator, relying on cudaMallocAsync and cudaFreeAsync, can remap physical pages when running of GPU memory, thus is more resistant to memory fragmentation.

Describe the solution you'd like
Set the async allocator as the default.

Describe alternatives you've considered
Continue improving the arena allocator is possible, but since it doesn't remap memory, probably won't ever be as fragmentation resistance as the async allocator.

Additional context
Set this at the beginning of 22.04 for more testing.

[rongou]

Since cudaMallocAsync requires CUDA 11.2, we'll do a version check and only switch to the async allocator if the cuda driver is above 11.2.