[WIP] Reduce peak memory to 8 GB

[andy-yang-1]

When we apply delta, we load two complete models at the same time, which puts a lot of strain on the CPU memory. This PR allows us to apply delta within 10GB of memory.

• run with

  python3 -m fastchat.model.apply_delta \
    --base /path/to/llama-13b \
    --target /output/path/to/vicuna-13b \
    --delta lmsys/vicuna-13b-delta-v0