Reduce memory consumption in b-tagging, by sharing NNs

This shouldn't change any outputs. It should just reduce memory consumption and startup time.

The GN2 transformer is getting kind of big: it pulls around 80 MB of memory per instance. Unfortunately we currently duplicate it a few times for various reasons: different jet collections with the same tagger, different "flip" configurations, etc. The idea is to introduce an optional service that lets us share the underlying NN where possible.

I also removed some very bad track decoration code which would break the EDM if anyone ever enabled it by accident.

Steps:

• Add an option redefinition constructor for the GNN
• Add a service that GNNTool can use to register the GNNs we've constructed so far
• Add the service in FlavorTagNNConfig.py and use it in our standard NN Cfg functions
• Remove super unsafe track decoration stuff from GNN (this should have been done much earlier)
• See if it works

Tagging @svanstro who might be mucking around some of the same code, e.g. in !71095 (merged).

PhysicsAnalysis/JetTagging/FlavorTagDiscriminants/FlavorTagDiscriminants/INNSharingSvc.h

@@ -15,10 +15,13 @@ namespace FlavorTagDiscriminants
   {
   public:
     DeclareInterfaceID(FlavorTagDiscriminants::INNSharingSvc, 1, 0);
-    bool contains(const std::string& nn_name);
-    void insert(const std::string& nn_name,
-                const GNNOptions& opts,
-                const GNN&);
+    virtual bool contains(const std::string& nn_name);
+    virtual void insert(const std::string& nn_name,
+                        const GNNOptions& opts,
+                        const std::shared_ptr<const GNN>&);
+    virtual std::shared_ptr<const GNN> at(
+      const std::string& nn_name,
+      const GNNOptions& opts);
   }
 }