nvmf/tcp: set pending_flush flag early

Before commit b402dea6 posix_sock_flush() handled zero copy first and
only then performed actual flush. That order mattered as part of the
_sock_check_zcopy() was calling req->cb_fn() in the
spdk_sock_request_complete().
NVMf TCP in one of possible flows has cb_fn set to
nvmf_tcp_pdu_c2h_data_complete which in turn leads to
spdk_sock_writev_async() in the _tcp_write_pdu().
If that happens then there is an outstanding msg in socket list that
requires flush.
Before b402dea6 it was flushed once req->cb_fn() ended so there was no
need to schedule additional flush. Now it is required.
pending_flush flag is still needed though to avoid the storm of flush
requests messages that might overflow the thread messages list.

Change-Id: Icb695a22775d3fbcab949702fe84dd07901acbce
Signed-off-by: Krzysztof Goreczny <krzysztof.goreczny@dell.com>
Suggested-by: Konrad Sztyber <ksztyber@nvidia.com>
Reviewed-on: https://review.spdk.io/c/spdk/spdk/+/26347


Community-CI: Mellanox Build Bot
Tested-by: SPDK Automated Test System <spdkbot@gmail.com>
Reviewed-by: Jim Harris <jim.harris@nvidia.com>
Reviewed-by: Konrad Sztyber <ksztyber@nvidia.com>