Fix chunk unloading regression from upstream merge

Commit a24318ee changed drop() to immediately remove chunks from cache,
bypassing the deferred m_toDrop save/unload pipeline. This caused missing
chunks on dedicated servers, iterator invalidation in dropAll() and
ServerLevel::save(), and entity duplication (item frames) from chunks
being reloaded without their entities first being removed from the level.

- ServerChunkCache::drop(): restore m_toDrop queue instead of immediate
  cache removal, so tick() can save/unload/move to unloadedCache safely
- MultiPlayerChunkCache::drop(): restore soft-unload (keep chunk in cache
  with loaded=true) instead of nulling cache and hasData
- PlayerChunkMap::setRadius(): remove dropAll() call when reducing radius,
  the per-chunk removal loop already handles out-of-range chunks

Minecraft.Client/PlayerChunkMap.cpp

@@ -824,11 +824,6 @@ void PlayerChunkMap::setRadius(int newRadius)
 			}
 		}
 
-		if (newRadius < radius)
-		{
-			level->cache->dropAll();
-		}
-
 		assert(radius <= MAX_VIEW_DISTANCE);
 		assert(radius >= MIN_VIEW_DISTANCE);
 		this->radius = newRadius;