Add realloc based `TexelBuffer<T>::into_vec(self) -> Vec<T>`

[197g]

This can be more efficient than use to_owned() on a deref'd slice. By letting the allocator decide if the allocation can be reused with a different alignment, it can elide a memcpy operation in some cases. Presumably, at least a lot of cases. Other alignment-aware allocators may be able to just move the allocation to a new bucket instead.