[Algs] Run sparse communication on a single large allocation

[tdavidcl]

Payload should be replaced by something like:

struct MPI_Message {
    sham::DeviceBuffer<u8> & bytebuf; // maybe use a host/device variant
    u64 offset, length;
    i32 rank_sender, rank_receiver;
}

where bytebuf is shared by multiple messages. Additionally the direct GPU toggle should be handled manually where the allocation is swapped by a host one last minute if gpu not avail.