Inspired by #2314, #2310, #2300, and others, I built mlpack with -fsanitize=address and ran mlpack_test. The results were... noisy. But I took some time and I fixed everything (except for one, detailed later).

• All of the ANN layers are properly freed now; it took a while to get this write, and DeleteVisitor will now delete internally-held layers when the layer has a Model() method. This may solve some issues people were having.

• CNN layers don't make an alias of the input from Forward() anymore. This should fix Can't Train a Normal CNN. #2146 .

• Test code properly frees things that it allocates. There were some tricky places where deallocation had been overlooked.

• Bindings handle exceptions properly and deallocate any allocated memory when an exception is thrown. It makes the code a bit uglier, but it's safer.

There is one bug I didn't solve, which surfaces in the reinforcement learning code, in q_learning_impl.hpp:

     // Update target network
     if (totalSteps % config.TargetNetworkSyncInterval() == 0)
      targetNetwork = learningNetwork;

This code actually copies an FFN, and each individual layer will require a copy constructor to fix that memory leak. That's the same issue as in #2314, and anyway, that isn't solved here. But everything else in the tests is.

I do think it's important that we merge this before a release, otherwise we'll probably get lots of segfault and invalid memory access reports. :)