We must not update cached tail in uct_mm_ep_flush() if there are any
pending elements. We may get new send resources but not use them, so
flush could return UCS_OK while there pending requests.

Fixes openucx#3052

- Ignore remote completions in case of forced flush (CLOSE_MODE_CANCEL)

- The UCP_EP_FLAG_FLUSH_STATE_VALID flag can't be used because it's not
  valid in release mode. Instead, use the UCP_EP_FLAG_DEST_EP flag as an
  indirect indication that we may have some operations which wait for
  software rma/amo completion.

- Fix missing initializion of flush state in case of client/server
  connection establishment with p2p lanes and *without* pre-request.