Reduce the peak memory requirement when applying delta

[zhisbug]

Seeing many complaints about the required peak memory being too high. We would like to keep the peak memory of the following command be less than 8GB.

python3 -m fastchat.model.apply_delta \
    --base /path/to/llama-13b \
    --target /output/path/to/vicuna-13b \
    --delta lmsys/vicuna-13b-delta-v1.1

[andy-yang-1]

I am working on reducing the peak memory, and the peak memory now is 11372.2 MiB. See #402 and we can further improve it.

[zhisbug]

@andy-yang-1 Great work, thanks!