support qwen2 hf<->mcore ckpt converter

[wenyujin333]

usage example:

TARGET_EP_SIZE=${TARGET_EP_SIZE:-1}
TARGET_PP_SIZE=${TARGET_PP_SIZE:-1}
TARGET_CKPT_FORMAT=${TARGET_CKPT_FORMAT:-"torch_dist"}

torchrun --nproc-per-node=1 --nnodes=1 checkpoint/convert.py \
--model-type GPT \
--loader qwen2_hf \
--saver mcore \
--target-tensor-parallel-size ${TARGET_TP_SIZE} \
--target-pipeline-parallel-size ${TARGET_PP_SIZE} \
--target-expert-parallel-size ${TARGET_EP_SIZE} \
--load-dir ${HF_FORMAT_DIR} \
--save-dir ${MCORE_FORMAT_DIR} \
--tokenizer-model ${TOKENIZER_MODEL} \
--target-ckpt-format ${TARGET_CKPT_FORMAT}```