Fix subflake handling

Relative 'path:' flake inputs now use the containing flake's
InputAccessor. This has the following implications:

* They're no longer locked in the lock file.

* They don't cause an additional copy to the store.

* They can reference the containing directory (i.e. a subflake can
  have an input '../foo', so long as it doesn't go outside the
  top-level containing flake).

Note: this is not a complete fix for subflake handling, since the lock
file currently makes it ambiguous what the containing flake is. We'll
probably need to add another field to the lock file for that.

Fixes #6352.

src/libfetchers/fetchers.hh

@@ -63,6 +63,8 @@ public:
        one that contains a commit hash or content hash. */
     bool isLocked() const { return locked; }
 
+    bool isRelative() const;
+
     bool hasAllInfo() const;
 
     bool operator ==(const Input & other) const;
@@ -151,6 +153,9 @@ struct InputScheme
     {
         throw UnimplementedError("getAccessor");
     }
+
+    virtual bool isRelative(const Input & input) const
+    { return false; }
 };
 
 void registerInputScheme(std::shared_ptr<InputScheme> && fetcher);