Added ExposedModel to produce predictions

[MatsMoll]

The main feature was to add a way to define exposed models.

This makes it easier to find out where a model contract can be used, and how to use the model.
Therefore, the main purpose is to answer the following questions:

• Where are our models located? An API, model registry, etc?
• Which kind of format do we expect to provide in order to use the model? JSON column wise, row wise, protobuf?

@model_contract(
    input_features=[...],
    exposed_model=ExposedModel.in_memory_mlflow(
        model_name="taxi_eta",
        model_alias="Champion",

        prediction_column="predicted_eta",
        predicted_at_column="predicted_at",
        model_version_column="model_version",
        model_contract_version_tag=None  # Since there are no versions, will this mean the literal array
    )
)
class TaxiRegressionModel:
    trip_id = String().as_entity()
    predicted_eta = Float()
    predicted_at = EventTimestamp()
    model_version = String().as_model_version()

However, this also means that we can potentially use the modells with the following:

store = await FeatureStore.from_dir()
preds = await store.model("taxi_eta").predict_over({
    "trip_id": [...],
}).to_polars()

Or store them in the output source directly?

store = await FeatureStore.from_dir()
await store.model("taxi_eta").predict_over({
    "trip_id": [...],
}).upsert_into_output_source()

This also lead to the development of some generated Ollama contracts:

from aligned.exposed_models.ollama import ollama_classification_contract, ollama_embedding_contract

Classification = ollama_classification_contract(
    contract_name="some_classification",
    inputs=[features.x, features.y, features.z],
    prompt_template="You are ... you have {x}, {y}, and {z}, is it true or false?",
    ground_truth=features.a,
    entities=[features.row_id],
    output_source=FileSource.parquet(...),
    endpoint="ollama_endpoint",
)

Other improvements:

• Added a loaded_columns to not load more data than needed.