Store children vectors in a separate data collection.

This reduces per-zone memory cost by 9 bytes if there are no children
and increases it by 4 bytes, if there are children. This is universally
a better solution, as the following data shows:

+++ /home/wolf/desktop/tracy-old/android.tracy +++
Vectors: 2794480
Size 0: 2373070 (84.92%)
Size 1: 70237 (2.51%)
Size 2+: 351173 (12.57%)
+++ /home/wolf/desktop/tracy-old/asset-new.tracy +++
Vectors: 1799227
Size 0: 1482691 (82.41%)
Size 1: 93272 (5.18%)
Size 2+: 223264 (12.41%)
+++ /home/wolf/desktop/tracy-old/asset-new-id.tracy +++
Vectors: 1977996
Size 0: 1640817 (82.95%)
Size 1: 97198 (4.91%)
Size 2+: 239981 (12.13%)
+++ /home/wolf/desktop/tracy-old/asset-old.tracy +++
Vectors: 1782395
Size 0: 1471437 (82.55%)
Size 1: 88813 (4.98%)
Size 2+: 222145 (12.46%)
+++ /home/wolf/desktop/tracy-old/big.tracy +++
Vectors: 180794047
Size 0: 172696094 (95.52%)
Size 1: 2799772 (1.55%)
Size 2+: 5298181 (2.93%)
+++ /home/wolf/desktop/tracy-old/darkrl.tracy +++
Vectors: 12014129
Size 0: 11611324 (96.65%)
Size 1: 134980 (1.12%)
Size 2+: 267825 (2.23%)
+++ /home/wolf/desktop/tracy-old/mem.tracy +++
Vectors: 383097
Size 0: 321932 (84.03%)
Size 1: 854 (0.22%)
Size 2+: 60311 (15.74%)
+++ /home/wolf/desktop/tracy-old/new.tracy +++
Vectors: 77536
Size 0: 63035 (81.30%)
Size 1: 8886 (11.46%)
Size 2+: 5615 (7.24%)
+++ /home/wolf/desktop/tracy-old/selfprofile.tracy +++
Vectors: 22940871
Size 0: 22704868 (98.97%)
Size 1: 73000 (0.32%)
Size 2+: 163003 (0.71%)
+++ /home/wolf/desktop/tracy-old/tbrowser.tracy +++
Vectors: 962682
Size 0: 695380 (72.23%)
Size 1: 43007 (4.47%)
Size 2+: 224295 (23.30%)
+++ /home/wolf/desktop/tracy-old/virtualfile_hc.tracy +++
Vectors: 529170
Size 0: 449386 (84.92%)
Size 1: 15694 (2.97%)
Size 2+: 64090 (12.11%)
+++ /home/wolf/desktop/tracy-old/zfile_hc.tracy +++
Vectors: 264849
Size 0: 220589 (83.29%)
Size 1: 9386 (3.54%)
Size 2+: 34874 (13.17%)

server/TracyWorker.hpp

@@ -111,6 +111,8 @@ private:
         flat_hash_map<uint64_t, uint16_t, nohash<uint64_t>> threadMap;
         Vector<uint64_t> threadExpand;
         std::pair<uint64_t, uint16_t> threadLast;
+
+        std::vector<Vector<ZoneEvent*>> m_zoneChildren;
     };
 
     struct MbpsBlock
@@ -173,7 +175,7 @@ public:
     // GetZoneEnd() will try to infer the end time by looking at child zones (parent zone can't end
     //     before its children have ended).
     // GetZoneEndDirect() will only return zone's direct timing data, without looking at children.
-    static int64_t GetZoneEnd( const ZoneEvent& ev );
+    int64_t GetZoneEnd( const ZoneEvent& ev );
     static int64_t GetZoneEnd( const GpuEvent& ev );
     static tracy_force_inline int64_t GetZoneEndDirect( const ZoneEvent& ev ) { return ev.end >= 0 ? ev.end : ev.start; }
     static tracy_force_inline int64_t GetZoneEndDirect( const GpuEvent& ev ) { return ev.gpuEnd >= 0 ? ev.gpuEnd : ev.gpuStart; }
@@ -189,6 +191,8 @@ public:
     const char* GetZoneName( const GpuEvent& ev ) const;
     const char* GetZoneName( const GpuEvent& ev, const SourceLocation& srcloc ) const;
 
+    tracy_force_inline const Vector<ZoneEvent*>& GetZoneChildren( int32_t idx ) const { return m_data.m_zoneChildren[idx]; }
+
     std::vector<int32_t> GetMatchingSourceLocation( const char* query ) const;
 
 #ifndef TRACY_NO_STATISTICS
@@ -294,8 +298,8 @@ private:
     uint16_t CompressThreadReal( uint64_t thread );
     uint16_t CompressThreadNew( uint64_t thread );
 
-    tracy_force_inline void ReadTimeline( FileRead& f, Vector<ZoneEvent*>& vec, uint16_t thread );
-    tracy_force_inline void ReadTimelinePre033( FileRead& f, Vector<ZoneEvent*>& vec, uint16_t thread, int fileVer );
+    tracy_force_inline void ReadTimeline( FileRead& f, ZoneEvent* zone, uint16_t thread );
+    tracy_force_inline void ReadTimelinePre033( FileRead& f, ZoneEvent* zone, uint16_t thread, int fileVer );
     tracy_force_inline void ReadTimeline( FileRead& f, Vector<GpuEvent*>& vec );
     tracy_force_inline void ReadTimelinePre032( FileRead& f, Vector<GpuEvent*>& vec );