Coming from the embedded world, I am far more used to the style bluetomcat showed where the caller handles allocation. It takes more work to use, and you have to make sure to not use it before initializing or after destroying, as you said. The advantage is that the caller has full control of memory allocation. The object can be a local variable, a static, malloc:ed, come from a custom allocator (say a slab allocator), or be a part of a larger struct.

In games and embedded systems, a very common pattern is batched heap allocation: a single allocation of an array of foo, and/or an array of foo mixed with other types, or sometimes a memory managed data structure like a memory pool. This is one big reason that the C++ Standard Template Library was shunned in game studios for a long time; it automatically did it’s own heap allocations, sometimes at inopportune times IIRC like in constructors, and didn’t let the caller choose. EA wrote their own version (EASTL) that was game friendly because it allowed control of the allocator.

In GPU programming, there are actual different memory sub-systems, with different sizes and different performance implications, so it’s critical that the caller is able to do the allocation & deallocation any way they want. This is why most well designed GPU APIs rarely allocate GPU memory inside the API, but instead are designed to work with caller-provided pointers to buffers.