Enable preparing nested OutputSchemas for serialization

[dbogunowicz]

Currently, the prep_outputs_for_serialization function assumes that pipeline_outputs cannot be arbitrarily nested (allows up to one level of nesting). This is why, for more complex BaseModels returned by the python (e.g. TextGenerationOutput), the function will not properly convert nested numpy arrays to lists for serialization.

My diff makes sure that any pipeline_output: BaseModel that contains an arbitrary number and nesting depth of any fields that are either a BaseModel, list of numpy.ndarray is supported.

As a result, we can now serialize logits of the LLM output, that previously was not properly converted from numpy array to a list.

import requests

model_path = "hf:mgoin/TinyStories-1M-deepsparse"
prompt =  ["name one former president of the USA"]
server_address = "http://0.0.0.0:5543/v2/models/gen/infer"

payload = {"prompt": prompt, "output_scores": True, "include_prompt_logits": True}
response = requests.post(server_address, json=payload)
response = response.json()

num_tokens_prompt_and_generated = len(response["generations"][0]["score"])

payload = {"prompt": prompt, "output_scores": True}
response = requests.post(server_address, json=payload)
response = response.json()

num_tokens_generated = len(response["generations"][0]["score"])
num_tokens_prompt = num_tokens_prompt_and_generated - num_tokens_generated
print(f"Number of prompt tokens: {num_tokens_prompt}")
print(f"Number of generated tokens: {num_tokens_generated}")

returns

Number of prompt tokens: 6
Number of generated tokens: 3