Allow Passing in Instantiated Teacher

[Satrat]

Asana ticket: https://app.asana.com/0/1206109050183159/1206788571042422/f

We had missed support for passing in an instantiated teacher to the finetuning code, this PR adds in it

Testing

script would fail on main, training starts as intended here:

from sparseml.transformers import compress, SparseAutoModelForCausalLM, SparseAutoTokenizer, load_dataset

model = SparseAutoModelForCausalLM.from_pretrained("zoo:llama2-7b-ultrachat200k_llama2_pretrain-base", device_map="auto")
teacher = SparseAutoModelForCausalLM.from_pretrained("zoo:llama2-7b-open_platypus_orca_llama2_pretrain-base", device_map="auto")
tokenizer = SparseAutoTokenizer.from_pretrained("zoo:llama2-7b-ultrachat200k_llama2_pretrain-base")
recipe = "zoo:llama2-7b-ultrachat200k_llama2_pretrain-pruned40"
dataset = load_dataset("garage-bAInd/Open-Platypus")

def format_data(data):
    data["text"] = data["instruction"] + data["output"]
    return data

dataset = dataset.map(format_data)

compress(
    model=model,
    tokenizer=tokenizer,
    distill_teacher=teacher,
    dataset=dataset,
    recipe=recipe,
    num_train_epochs=1,
    output_dir="./output",
)