Revamp of rmw_context_impl_s

[clalancette]

This PR is an attempt to fix two particular issues surrounding rmw_context_impl_s:

1. It could be the case that a graph_sub_data_handler call could attempt to access a Data pointer that had already been freed.
2. Sometimes on shutdown, we would end up in a deadlock situation where graph_sub_data_handler was being called by Zenoh while Data::shutdown was trying to destroy the session. This would end up in an AB/BA deadlock.

It's probably best to review this PR patch-by-patch, as this will show you how I got to this solution, but in short:

1. We first make rmw_context_impl_s::Data private within rmw_context_impl_s.cpp. This isn't strictly required, but it does form more of a PIMPL pattern and is tidier.
2. We remove the unnecessary Data::subscribe_to_ros_graph method and just put the code inline.
3. We remove the unused Data::is_initialized_ member.
4. We remove the unused Data::allocator_ member.
5. We stop storing Data::liveliness_str_; we only need it during the constructor.
6. We mark the Data structure as final, and remove enable_shared_from_this (we don't currently need it).
7. We keep a global map of Data structure pointers to Data structure shared_ptr. We give the Data structure pointer as a "key" to graph_sub_data_handler, but it is strictly used as a key to look up the shared_ptr. Once we have the shared_ptr, we are guaranteed that the structure will not be deleted while we are using it.
8. Possibly most controversially, we move all work on the Data members within Data, mark Data as a class (rather than a structure), and mark all of the member variables private. This makes it crystal-clear which entity is responsible for which. With this in place, rmw_context_impl_s is a thin wrapper over Data, that really only passes through calls and manages the addition and removal of Data from the global map.

With all of this in place, I can no longer reproduce the crash that I was seeing locally. Also, with this and with ros2/rclpy#1353 in place, I am down to a single reproducible failing test across [rcl, rcl_action, rcl_lifecycle, rclcpp, rclcpp_action, rclcpp_lifecycle, rclpy, test_rclcpp, test_communication].

[ahcorde]

I added in this comment a summary of test failures.

[Yadunund]

Thanks for cleaning this up! Left some feedback to consider!

[Yadunund]

Nit: For better readability, can we first define Data and all its members here first before we implement them later in this file?

///=============================================================================
class rmw_context_impl_s::Data
{
public:
  Data(...);

  rmw_ret_t shutdown();
  ...
private:
  z_owned_session_t session_;
  ...
};

///=============================================================================
rmw_context_impl_s::Data::Data(...)
{
...
}

That should also minimize the diff.

[clalancette]

Yeah, we can do that. I generally find that unnecessarily verbose, because you are doing the same thing but with more lines.

That said, I don't feel super strongly about it. Let me know if you really prefer that style and I can switch to it.

[Yadunund]

Not a strong requirement from my end