[Eager, Performance Optimization] modify AllocateFrom to reduce deconstruction...

[Eager, Performance Optimization] modify AllocateFrom to reduce deconstruction of shared_ptr (#48548)

paddle/phi/core/dense_tensor.cc

@@ -128,7 +128,16 @@ void* DenseTensor::AllocateFrom(Allocator* allocator,
   if (!holder_ || holder_->size() < bytes + meta_.offset) {
     meta_.offset = 0;
     VLOG(10) << "Allocate data with bytes: " << bytes;
-    ResetHolder(allocator->Allocate(bytes));
+    auto holder = allocator->Allocate(bytes);
+    if (holder_) {
+      PADDLE_ENFORCE_LE(
+          numel() * static_cast<int64_t>(SizeOf(dtype)) +
+              static_cast<int64_t>(meta_.offset),
+          static_cast<int64_t>(holder->size()),
+          phi::errors::InvalidArgument(
+              "The size of Holder is not enough to store the Tensor."));
+    }
+    holder_ = std::move(holder);
   }
 
   return reinterpret_cast<void*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +