perf(server): avoid unnecessary reparses and cache navigation definitions

Introduce several changes to reduce full-document reparses, lock contention and
redundant work when handling TclOO analysis and navigation:

- Add a cheap may_contain_classes pre-check and several cursor/receiver
  heuristics so completions, signature help and tcloo definitions skip the
  expensive marker reparse when the document cannot contain useful OO info.
- Allow passing an existing parsed tree into tcloo completion/signature/definition
  helpers; update callers to use the server's cached tree when available.
- Use a thread-local parser for request-time parse_source to avoid blocking the
  shared parser during background indexing, and add navigation_state() which
  returns cached definition identities (invalidated on index generation changes).
- Add a cheap name pre-filter (_may_resolve_to) for symbol matching and only
  run class highlighting when classes may exist.

These changes reduce contention and repeated parsing, improve responsiveness for
requests during background indexing, and cache navigation definition identities.
Tests were added/updated to assert caching and non-blocking behavior.

server/src/tools/inlay_hint.py

@@ -200,7 +200,7 @@ class InlayHintGenerator(Visitor):
 
         walk(tree)
         if self.parameter_names != "none":
-            for call in resolved_method_calls(self.source, self.external_classes):
+            for call in resolved_method_calls(self.source, self.external_classes, tree):
                 if not self._node_intersects_requested_range(call.command):
                     continue
                 parameters = method_parameters(call.parameters)