Incorrect result report for UCP tag_bw test

[SeyedMir]

Describe the bug

The UCP tag_bw test reports the perf values from the client side, which leads to incorrect and misleading bandwidth values. The results should instead be reported by the server process because the client only waits for the local completion of the send operations.

Steps to Reproduce

On a single node with NVLink

• ucx_perftest -f -t tag_bw -s 128000000 -m cuda,host &
• ucx_perftest -f -t tag_bw -s 128000000 -m cuda,host localhost
• UCX version: Git branch 'master', revision 21c1594
• Env vars: UCX_RNDV_FRAG_MEM_TYPE=cuda UCX_TLS=tcp,cuda_copy,cuda_ipc,self