provide a service to long-running GPU kernels to flush incoming RDMA traffic

[drossetti]

For instance this is needed when a kernel is waiting for incoming RDMA traffic like data + flag, by polling on the flag. In that case, the GPU memory model does not guarantee automatic consistency for GPU engines.