Bugfix: Remove default tensor board writer which may cause permission problem

P.S. If you have problem like "permission denied" of "/home/admin", please try the new fixed release v0.4.3

Support more open source models like Qwen2, Qwen2-moe, Starcoder2, etc.
Support QLoRA + Deepspeed ZeRO3 / FSDP, which is efficient for very large models.

Updates:

1. Mainly forMFTCoder-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.

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 theATorch frameworks, which is a fast distributed training framework of LLM.