Handle sparse copy domains for GPU indirect copies

[apryakhin]

That includes a number of optimizations we need to get inside for indirect copies such as efficient handling of sparse copy domains, gathering from multiple buffers over NVLink, handling out-of-range accesses.

• realm: add cuda-gather-dma refactoring #326