The use of `shared_merge_buffer` should depend on the tensor node's `memory_mode` (on all backends)

[lukstafi]

Currently this is very broken and confused, e.g. the cuda backend always uses it when merging by-copy.