Final merge still needs to be loaded into memory

[one-some]

currently, the entire final merge still needs to be loaded into memory, be it gpu, disk, or cpu. theoretically, we can calculate the final weight values while only ever loading n weights at a time (merge targets).

torch.save/pickle seems to be a headache for that sort of thing (see the maybe-0b branch), unless we do shards or something. safetensors seeeeeeeeems more approachable for something like this as we (hopefully) wont have to muck about in pickle internals anymore.