Feature Request: Add TimeStral as a default time-series prediction engine

[ottlukas]

I'd like to propose integrating TimeStral, Mistral AI's family of open-weight time-series foundation models, as a default, zero-shot prediction engine in StreamPipes — similar to how StreamPipes ships ready-to-use ML pipeline elements today.
Why TimeStral is a particularly good fit for StreamPipes:

• Zero-shot forecasting: forecasts unseen series with no training on user data — ideal for newly installed machines/sensors with little history.
• Probabilistic & multivariate: 21 quantiles per step; supports multiple related series plus past-only and known-future covariates (numeric and categorical).
• CPU-friendly: PyTorch inference with no GPU needed (5M model ≈ 10 ms, 209M model ≈ 140 ms per forecast on a server CPU) — practical for edge deployments, which is a key StreamPipes use case.
• Long context: up to 32,768 historical time steps; horizons up to 1,024 steps per forward pass.
• Open weights (Modified MIT) in four sizes (5M / 28M / 123M / 209M) on Hugging Face, so it can ship or be downloaded without licensing friction and runs fully on-prem, even air-gapped.
• State-of-the-art accuracy: the 209M model ranks first on GIFT-Eval, TIME, and fev-bench with only a fraction of the parameters of comparable models.