Remove NCCL_NET_GDR_LEVEL=LOC from VERL/OpenRLHF Dockerfiles (#1106)

NCCL_NET_GDR_LEVEL=LOC disables GPU Direct RDMA over the network — every
cross-node GPU↔GPU transfer is staged through host memory, defeating the
purpose of EFA on p5/p5e/p5en/p6.

Removing the override lets the aws-ofi-nccl plugin (bundled with the EFA
installer, built with --enable-platform-aws) select the appropriate GDR
level via the per-instance NCCL_TOPO_FILE it auto-loads.

Author: mvinci12

3.test_cases/pytorch/openrlhf/Dockerfile

@@ -63,7 +63,6 @@ ENV FI_EFA_USE_DEVICE_RDMA=1
 ENV FI_EFA_FORK_SAFE=1
 ENV FI_EFA_ENABLE_SHM_TRANSFER=1
 ENV NCCL_PROTO=simple
-ENV NCCL_NET_GDR_LEVEL=LOC
 ENV NCCL_SOCKET_IFNAME=^docker,lo,veth
 ENV NCCL_TUNER_PLUGIN=/opt/amazon/ofi-nccl/lib/x86_64-linux-gnu/libnccl-ofi-tuner.so
 ENV PMIX_MCA_gds=hash

3.test_cases/pytorch/verl/hyperpod-eks/rlvr/Dockerfile

@@ -78,7 +78,6 @@ ENV FI_EFA_USE_DEVICE_RDMA=1
 ENV FI_EFA_FORK_SAFE=1
 ENV FI_EFA_ENABLE_SHM_TRANSFER=1
 ENV NCCL_PROTO=simple
-ENV NCCL_NET_GDR_LEVEL=LOC
 ENV NCCL_SOCKET_IFNAME=^docker,lo,veth
 ENV NCCL_TUNER_PLUGIN=/opt/amazon/ofi-nccl/lib/x86_64-linux-gnu/libnccl-ofi-tuner.so
 ENV PMIX_MCA_gds=hash

3.test_cases/pytorch/verl/kubernetes/rlvr/Dockerfile

@@ -79,7 +79,6 @@ ENV FI_EFA_USE_DEVICE_RDMA=1
 ENV FI_EFA_FORK_SAFE=1
 ENV FI_EFA_ENABLE_SHM_TRANSFER=1
 ENV NCCL_PROTO=simple
-ENV NCCL_NET_GDR_LEVEL=LOC
 ENV NCCL_SOCKET_IFNAME=^docker,lo,veth
 ENV NCCL_TUNER_PLUGIN=/opt/amazon/ofi-nccl/lib/x86_64-linux-gnu/libnccl-ofi-tuner.so
 ENV PMIX_MCA_gds=hash