[AMD] job.slurm: inject MoRI-IO QP knobs via docker -e for vllm-disagg

env.sh exports the MORI_IO_* QP knobs but they do not propagate to the vLLM
worker processes, so inject them into the container base env via docker -e.
MORI_IO_TC_DISABLE is intentionally omitted: the TC value is detected per-node
in env.sh and cannot reach the workers, so enabling TC steering without a TC
value would just fall back to TC=0; leave MoRI-IO at its library default instead.

Author: Duyi-Wang

benchmarks/multi_node/amd_utils/job.slurm

@@ -403,6 +403,10 @@ if [[ "$ENGINE" == "vllm-disagg" ]]; then
         -e PROXY_STREAM_IDLE_TIMEOUT=\${PROXY_STREAM_IDLE_TIMEOUT:-300}
         -e VLLM_MORIIO_CONNECTOR_READ_MODE=\${VLLM_MORIIO_CONNECTOR_READ_MODE:-1}
         -e PYTHONPYCACHEPREFIX=/tmp/pycache
+        -e MORI_IO_SQ_BACKOFF_TIMEOUT_US=\${MORI_IO_SQ_BACKOFF_TIMEOUT_US:-50000}
+        -e MORI_IO_QP_MAX_SEND_WR=\${MORI_IO_QP_MAX_SEND_WR:-16384}
+        -e MORI_IO_QP_MAX_CQE=\${MORI_IO_QP_MAX_CQE:-32768}
+        -e MORI_IO_QP_MAX_SGE=\${MORI_IO_QP_MAX_SGE:-2}
     )
 elif [[ "$ENGINE" == "atom-disagg" ]]; then
     DOCKER_ENV_ENGINE=(