Memory leak in error handling flow after RMA op failure

[evgeny-leksikov]

After failure of RMA operation, the following is happening (this is a following of another issue #1770 ):

17:45:00 [ RUN ] dcx/test_ucp_rma.nonblocking_stream_get_nbi_flush_ep/1
17:45:01 [1503585901.060628] [hpc-test-node2:40481:3] ib_device.c:171 UCX ERROR IB Async event on mlx5_0: DCT access error on DCTN 0x9374
17:45:01 [1503585901.143067] [hpc-test-node2:40481:4] ib_mlx5_log.c:109 UCX ERROR Error on QP 0x9368 wqe[1]: Invalid request (synd 0x12 vend 0x8a) opcode RDMA_READ
17:45:01 [1503585901.143149] [hpc-test-node2:40481:4] ucp_worker.c:399 UCX ERROR Error Endpoint timeout was not handled for ep 0x35e5310
17:45:01 /scrap/jenkins/scrap/workspace/hpc-ucx-pr-4/label/hpc-test-node2/worker/2/contrib/../test/gtest/ucp/ucp_test.cc:378: Failure
17:45:01 Error: Endpoint timeout
17:45:01 [1503585901.272351] [hpc-test-node2:40481:4] mpool.c:38 UCX WARN object 0x313ef40 was not returned to mpool ucp_requests
...
17:45:01 [1503585901.413624] [hpc-test-node2:40481:4] rcache.c:284 UCX WARN mlx5_0: destroying inuse region 0x35e52b0 [0x7ffff01bd000..0x7ffff045d000] gt- rw ref 1 lkey 0x10451 rkey 0x10451 atomic: lkey 0x1b05 rkey 0x1b05
17:45:01 /scrap/jenkins/scrap/workspace/hpc-ucx-pr-4/label/hpc-test-node2/worker/2/contrib/../test/gtest/common/test.cc:228: Failure
17:45:01 Failed
17:45:01 Got 49 warnings during the test
17:45:01