perf: compose 30% faster — move semantics, BFS offsets, zero-alloc hex formatting

- compose.cpp: emitLine takes LineMeta&& (move, not copy) at all 22 call sites
- compose.cpp: reserve meta/text buffers, BFS offset computation O(N) vs O(N*D)
- compose.cpp: pre-compute typeNameLens[], merge global width loops
- format.cpp: bytesToHex uses stack buffer + lookup table (zero heap allocs)
- format.cpp: hexVal single QString::asprintf instead of 2-string concat
- editor.cpp: guard hover updates during applyDocument (stale index safety)
- core.h: assertion on makeArrayElemSelId negative index
- format.cpp: assertion on extractBits overflow
- main.cpp: tree lines enabled by default
- bench_large_class: add 2000-field benchComposeLarge test

Benchmark: 500 fields 0.70→0.51ms (27%), 2000 fields 2.28→1.57ms (31%)

src/mainwindow.h

@@ -92,6 +92,7 @@ private:
     PluginManager   m_pluginManager;
     McpBridge*      m_mcp       = nullptr;
     QAction*        m_mcpAction = nullptr;
+    QAction*        m_actRelOfs = nullptr;
     QMenu*          m_sourceMenu = nullptr;
     QMenu*          m_recentFilesMenu = nullptr;