fix(base): delete outer Communicator in CommDestroy

[GordonYang1]

Summary

Fixes the Communicator lifetime leak noted in the "Known Issues & Future Work" section of #10.

CommInitAll::Execute allocates the outer Communicator via new, but the previous CommDestroy only tore down the backend instance and never deleted the outer object. Every infiniCommInitAll / infiniCommDestroy pair leaked one Communicator.

Changes

• src/base/comm_destroy.h:
  • Narrow Execute to a concrete void *comm_handle signature, mirroring CommInitAll::Execute's void **comm_handle.
  • Add a null-handle check, consistent with CommInitAll.
  • Delete the Communicator only after the backend Apply returns kSuccess.

Backend implementations are intentionally untouched: ownership of the outer Communicator now lives entirely in the base layer, symmetric with the allocation in CommInitAll. Future backends such as NCCL, MCCL, HCCL, and others get correct lifetime handling without each backend having to remember to delete the outer object.

Test environment

Validated on a heterogeneous 2-node cluster with container-to-container direct connection over RDMA:

Node	Hardware	Container	IP
A	NVIDIA A100	iccl-nvidia	192.168.163.40
B	MetaX C550	iccl-metax-clean	192.168.162.49

• OS / container: --network host --ipc host --privileged, /dev/infiniband mounted on both sides.
• OpenMPI: 4.1.6 (/opt/openmpi-4.1.6), built with --with-ucx=/opt/ucx-1.17.0.
• UCX: 1.17.0 (/opt/ucx-1.17.0), built with --with-verbs --with-rdmacm.
• CUDA Toolkit on Node A; MACA SDK at /opt/maca on Node B.
• Inter-container SSH direct connection on port 22222, with no docker exec wrapper required for mpirun or icclrun --build.
• UCX transport configuration:
  • UCX_NET_DEVICES=mlx5_0:1
  • UCX_TLS=rc,rc_verbs,self,sm
  • UCX_RNDV_SCHEME=put_zcopy
• World size: 16 ranks, 8 NVIDIA + 8 MetaX.

Logs & Screenshots

all_reduce test (MetaX-NVIDIA heterogeneous)
all_reduce.log

all_gather test (MetaX-NVIDIA heterogeneous)
all_gather.log

reduce_scatter test (MetaX-NVIDIA heterogeneous)
reduce_scatter.log

broadcast test (MetaX-NVIDIA heterogeneous)
broadcast.log

all_to_all test (MetaX-NVIDIA heterogeneous)
all_to_all.log

[Ziminli]

同时麻烦 rebase 到最新，补充新增的示例程序运行日志文件。

[Ziminli]

麻烦修改一下 PR 标题的格式，可以参考其他已合入的 PR。