Re-instate memory limit #1670
Description
With the switch to the stream-ordered memory allocator, we lost the ability to limit the amount of GPU memory used by CUDA.jl. We filed this with NVIDIA as buf #3503815, but they mentioned that (awaiting proper support) it should be possible to emulate a limit by inspecting CU_MEMPOOL_ATTR_USED_MEM_CURRENT / CU_MEMPOOL_ATTR_RESERVED_MEM_CURRENT in our allocator.