Check for nullptr in TopicManager::ConnectPubToSub

[emersonknapp]

Fixes #2875

This was causing a very occasional crash problem in real runs, and is reliably reproducible by the method shown in #2875 (comment) - which case is explicitly forced by the added regression test. Though it's a pathological case, there is nothing illegal about the test's usage of a global, public API, and therefore it needs the protection.

The crash case of #2875 is particularly bad in that it allows the server process to be crashed by an external process' actions - all it takes is a mis-timed "sub" message to the server.

This issue also exists in Gazebo9 and will need to be backported.

[emersonknapp]

I'm not sure I'll be able to make the regression test reliably fail without the fix - it seems that the Boost assertion isn't doing anything in this test when playing around with it, though it does always crash my example program as mentioned in #2875.

I'm not sure exactly how to handle the testing for this - I could use input from someone more familiar with gazebo/boost test/build setup.

[j-rivero]

I have ported the test directly to gazebo11 branch, without the fixes in this PR, to check if the test is failing there:

• Mac
• Linux

[j-rivero]

I have ported the test directly to gazebo11 branch, without the fixes in this PR, to check if the test is failing there:

* Mac

* Linux

I see it crashing in both platforms and on my local machine so I think the test case is valid. After the patch the CI does not display this crashing anymore.

[j-rivero]

Thanks for the PR and iterations @emersonknapp