Improved IB host-no-atomic mode

[chhwang]

Fix potential memory inconsistency in IB host-no-atomic mode, and reduce latency overhead by introducing GDRCopy.

• IB no-atomic: 8-byte RDMA write-with-imm carries full 64-bit token to remote signal GPU buffer, which is read by the remote host before updating the inbound token for strict data-flag ordering.
• GDRCopy: recv thread reads token via BAR1 (CUDA) or uncached GPU memory (ROCm); no more cudaMemcpyAsync/CUDA stream