Lost service responses (#183, #74) (#187) (#209)

* Block rmw_send_response if response reader unknown

The client checks using rmw_service_server_is_available whether the
request it sends will be delivered to service, but that does not imply
that the (independent, as far as DDS is concerned) response reader of
the client has been discovered by the service.  Usually that will be the
case, but there is no guarantee.

Ideally DDS would offer an interface that allows checking the reverse
discovery, but that does not yet exist in either the specification or in
Cyclone.  This commit works around that by delaying publishing the
response until the number of request writers matches the number of
response readers.

Signed-off-by: Erik Boasson <eb@ilities.com>

* Change request headers to use rmw_request_id_t on the wire

Signed-off-by: Erik Boasson <eb@ilities.com>

* Precise check for matched client/service

Assign a unique identifier to each client/service on creation, add it
to the USER_DATA QoS of the reader and writer and use it for the request
ids.  This allows:

* rmw_service_server_is_available to only return true once it has
  discovered a reader/writer pair of a single service (rather than a
  reader from some service and a writer from some service); and
* rmw_send_response to block until it has discovered the requesting
  client's response reader and to abandon the operation when the client
  has disappeared.

The USER_DATA is formatted in the same manner as the participant
USER_DATA, this uses the keys "serviceid" and "clientid".

This is still but a workaround for having a mechanism in DDS to ensure
that the response reader has been discovered prior by the request writer
prior to sending the request.

Signed-off-by: Erik Boasson <eb@ilities.com>

* Address review comments

Signed-off-by: Erik Boasson <eb@ilities.com>

* Backwards compatibility

* Revert commit fb040c5 to retain the
  old wire representation;

* Embed the publication_handle of the request inside rmw_request_id_t,
  possible because reverting to the old wire representation frees up
  enough space, and use this in rmw_send_response to check for the
  presence of the client's reader;

* Clients and services without a client/service id in the reader/writer
  user data are treated as fully matched at all times.

* Replace ERROR by FAILURE to because of windows.h

Signed-off-by: Erik Boasson <eb@ilities.com>

* Timeout rmw_send_response after waiting 100ms for discovery

The discovery will eventually result in the client's reader being known
or its writer no longer being known, so a timeout is not necessary for
correctness.  However, if it ever were to block for a longish
time (which is possible in the face of network failures), returning a
timeout to the caller is expected to result in less confusion.

Signed-off-by: Erik Boasson <eb@ilities.com>

* Make iterators "const auto &"

Signed-off-by: Erik Boasson <eb@ilities.com>

* Add TODO for eliminating rmw_send_response blocking

Signed-off-by: Erik Boasson <eb@ilities.com>

Co-authored-by: eboasson <eb@ilities.com>