Add `InferenceModel` for safe production inference

[mkumar73]

Add an inference-only model interface for loading fitted DeepTab artifacts and running safe predictions in local, batch, or production environments.

InferenceModel should not replace model.predict(). The normal estimator API remains the main workflow for research and local experimentation. InferenceModel is for deployment use cases where the object should only support loading, validation, runtime inspection, and prediction.

Proposed API

from deeptab import InferenceModel

model = InferenceModel.load(
    "model.deeptab",
    device="cpu",
)

model.validate_input(X_new)
preds = model.predict(X_new)