junction-core: Fix an add-remove-add caching bug

[blinsay]

We managed to turn up an add-remove-add bug where the client cache was marking an xDS Cluster as not-found when removing it instead of just removing it from the cache. This was, unfortunately, fairly easy to turn up - speeding up ingest and changing route names in the smoke test (a6cb862) were enough to trigger it reliably.

The fix is trivial - on removing a resource from the cache, don't create a subscription node if one doesn't already exist.

     fn remove(&mut self, rtype: ResourceType, name: &str) {
-        if let Some(sub) = self.find_subcribed(rtype, name) {
+        if let Some(sub) = self.find(rtype, name) {
            self.reset_refs(sub);
          }
     }

For resources that are not wildcards, this was a noop, but for Clusters this would both tombstone resource data and create a wildcard subscription, which marks the resource as not-found.

Most of the fix is a connection level test I wrote trying to reproduce, and a cache-level test that scopes this more tightly. To repro, it was necessary to interleave a call to cache.collect between removing a reference to the cluster from the RouteConfig and removing the cluster itself from the cache.

More Debugging Tools

While reproducing, it was useful to have tracing enabled in Rust. Enabling tracing doesn't happen automatically, but there is now a hook that Python clients can call to enable the standard tracing subscriber. This isn't the best interface but is an improvement.