Thread-safe access to PublisherData

[Yadunund]

This PR replaces class rmw_publisher_data_t with class PublisherData which provides thread-safe access to publisher functions. An instance of PublisherData is created via the NodeData class. rmw_publisher->data is now set to rmw_node_t. The assumption here is that the node outlives the publisher similar to the assumption that the context outlives the node. From this rmw_node_t raw_ptr, we are able to obtain it's NodeData and then lookup the PublisherData for the publisher.

One nice thing this PR introduces is the mechanism where shutting down or destroying a parent entity, will shutdown the child entities. Eg. If rclcpp::shutdown() is invoked, it will first shutdown all the PublisherData, then NodeData, and finally the context. We also check for shutdown in the destructor. This is helpful when someone calls node.reset there by destroying an rclcpp::Node; this will now shutdown all child entities (only PublisherData so far) and then shutdown the node before clearing memory. We also check for is_shutdown before creating a publisher, publishing a message, etc. These principles will be extended to other entities in subsequent PRs.

[ahcorde]

just some mirmo things, but overall LGTM

[clalancette]

Because this is returning a raw pointer, this is probably not thread-safe.

That said, from what I can tell there is only one caller of this API, from rmw_get_gid_for_publisher. And there, what we are doing is copying the data into an out parameter.

That suggests to me that this function should be something like:

const void PublisherData::copy_gid(rmw_gid_t * gid) const
{
  std::lock_guard<std::mutex> lock(mutex_);
  memcpy(gid->data, gid_, RMW_GID_STORAGE_SIZE);
}

(and then update the caller down below as needed). That way we are doing the copy under the lock, which should be safe.

[Yadunund]

Good idea! Also had to do something similar for copying the GID into the attachment map 4edbee2