[feature] implement dcp for deepseek_v2

benchmark/kernels/all_reduce/bench_symm.sh

@@ -0,0 +1,25 @@
+: <<'result_in_h800'
+| msg_size   |   [AllGather] torch eager time |   [AllGather] pynccl symm graph time |   [ReduceScatter] pynccl eager time |   [ReduceScatter] pynccl symm graph time |
+|------------|--------------------------------|--------------------------------------|-------------------------------------|------------------------------------------|
+| 2.0 KiB    |                        33.7312 |                              2.88258 |                             34.1088 |                                  2.8155  |
+| 4.0 KiB    |                        27.4848 |                              2.97794 |                             29.6736 |                                  2.92332 |
+| 8.0 KiB    |                        25.6896 |                              3.43691 |                             25.8368 |                                  3.08607 |
+| 16.0 KiB   |                        60.4384 |                              4.94141 |                             25.28   |                                  3.14567 |
+| 32.0 KiB   |                        25.5264 |                              5.3824  |                             31.664  |                                  3.36464 |
+| 64.0 KiB   |                        26.8    |                              7.09412 |                             29.5456 |                                  3.61114 |
+result_in_h800
+
+export PER_NODE_GPU=8
+export WORLD_SIZE=1
+export RANK=0
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=12345
+export NCCL_GRAPH_MIXING_SUPPORT=0
+export NCCL_CUMEM_ENABLE=1
+export NCCL_NVLS_ENABLE=2
+
+torchrun  --nproc_per_node $PER_NODE_GPU \
+          --nnodes $WORLD_SIZE \
+          --node_rank $RANK \
+          --master_addr $MASTER_ADDR \
+          --master_port $MASTER_PORT benchmark_symm_mem.py