[core] Fix pre-existing extension point issue

[bmhowe23]

This PR fixes a pre-existing issue with our extension points. Without this change, when a new decoder plugin was added and cleanup_plugins was called at the end of a program, the extension points would sometimes crash. The solution is the following:

1. We must remove items from the registry to protect if someone were to dlclose a plugin library while keeping the parent libcudaq-qec.so library open.
2. Similar to Fixed double free error #327, we change the static variables to pointers rather than objects. This is actually the most important change.
3. We must use a mutex for protection against multi-threaded access to the registry. It must be a recursive_mutex to allow decoders to include other decoders (because the "get" method will try to lock multiple times in the call stack).

Also, a bit unrelated, but I also changed one of our tests to use a simulator that supports noise (stim) vs the default (which is often qpp depending on the environment). This eliminates thousands of warning messages that were introduced with NVIDIA/cuda-quantum#3113.