Add alpca json inference

[madhavajay]

The current inference is a little inflexible, I added something which lets you pass in json or a json file and either get json or a json file back out in the format of alpca style {"instruction":..."input":... "output"}.

I don't fully understand how the prompts are constructed so I think there are some changes there, but I have to test in the cloud due to no GPU on my Apple Silicon so it takes time to test.

For example:

$ accelerate launch scripts/finetune.py config-1b-qlora.yml --lora_model_dir="./qlora-output" --inference-json='{"input":"What is your name?"}' --inference-json-output=out.json

$ cat out.json
[{"input": "What is your name?", "output": "\"My Name Is... My name is Jessica, I am 32 years old.\""}]

[theobjectivedad]

Good morning, just an inform - this is similar to a feature in #291