GPU/TextureCache: Only use a single palette record for C16

Stops C16 textures larger than 256x256 being split up.

src/core/gpu_hw_texture_cache.cpp

@@ -1818,8 +1818,21 @@ void GPUTextureCache::UpdateVRAMWriteSources(VRAMWrite* entry, SourceKey source_
     return;
 
   // Add to the palette tracking list
-  auto iter = std::find_if(entry->palette_records.begin(), entry->palette_records.end(),
+  std::vector<VRAMWrite::PaletteRecord>::iterator iter;
+  if (source_key.HasPalette())
+  {
+    // Palette requires exact match.
+    iter = std::find_if(entry->palette_records.begin(), entry->palette_records.end(),
                         [&source_key](const auto& it) { return (it.key == source_key); });
+  }
+  else
+  {
+    // C16 only needs to match on the mode, palette is not used, and page doesn't matter.
+    // In theory we could extend the page skipping to palette textures too, but is it needed?
+    iter = std::find_if(entry->palette_records.begin(), entry->palette_records.end(),
+                        [&source_key](const auto& it) { return (it.key.mode == source_key.mode); });
+  }
+
   if (iter != entry->palette_records.end())
   {
     iter->rect = iter->rect.runion(write_intersection);

src/core/gpu_hw_texture_cache.h

@@ -58,7 +58,7 @@ struct TListNode
   TListNode<T>* next;
 };
 
-struct SourceKey
+struct alignas(4) SourceKey
 {
   u8 page;
   GPUTextureMode mode;