MFTCoder v0.3.0: Support more open source models, support Self-Paced Loss, support FSDP

Updates:

1. Mainly for MFTCoder-accelerate.
2. It now supports more open source models like Mistral, Mixtral(MoE), DeepSeek-coder, chatglm3.
3. It supports FSDP as an option.
4. It also supports Self-paced Loss as a solution for convergence balance in Multitask Fine-tuning.

v0.1.0 release: Multi Task Fintuning Framework for Multiple base modles

1. We released MFTCoder which supports finetuning Code Llama, Llama, Llama2, StarCoder, ChatGLM2, CodeGeeX2, Qwen, and GPT-NeoX models with LoRA/QLoRA.
2. mft_peft_hf is based on the HuggingFace Accelerate and deepspeed framework.
   mft_atorch is based on the ATorch frameworks, which is a fast distributed training framework of LLM.