Use latest peft/transformers/accelerate/bitsandbytes for 4-bit (qlora)

[arnocandel]

https://arxiv.org/abs/2305.14314
https://huggingface.co/blog/4bit-transformers-bitsandbytes

• update deps
• enable 4-bit generation
• enable 4-bit lora training

Addresses #136 more directly than #107, but only at runtime.

[arnocandel]

Testing 65B on 2x A6000Ada:
(env) arno@rippa:/nfs4/llm/h2ogpt(4bit)$ CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node=2 finetune.py --base_model=decapoda-research/llama-65b-hf --train_4bit=True --micro_batch_size=1 --run_id=3 --data_path=h2oai/openassistant_oasst1_h2ogpt_graded
works:
0%|▋ | 1/237 [04:27<17:33:37, 267.87s/it]
1%|█▎ | 2/237 [09:01<17:41:35, 271.04s/it]

[pseudotensor]

We will do more full testing of eval on 4bit vs. 8bit vs. 16bit. According to QLORA paper, as long as do LORA on all linear layers, then should be as good as 16bit. I'm unclear on speed of new Tim stuff.