ADD T0 (t0-alpha) zero-shot forecasting solver

[GeoffNN]

Adds theforecastingcompany/t0-alpha (~102M-param decoder-style patch transformer, Apache-2.0) as a local-inference forecasting solver, via the tfc-t0 package.

Design

• Forecasting only — the package exposes no embedding API, so classification/AD are skipped via skip().
• Native quantiles: requests the 5 levels the model was trained on (0.1, 0.25, 0.5, 0.75, 0.9) and returns them directly in ForecastOutput.
• Batched inference: T0 treats NaN as missing, so variable-length histories are left-padded with NaN into rectangular (B, C, T_max) batches — all (series, cutoff) windows grouped by channel count, chunked by a batch_size parameter to bound memory. Multivariate series go through the model's native [B, V, T] joint-forecast path.
• Model loading happens in set_objective (untimed), mirroring the Chronos-2 solver.

Testing

• Verified end-to-end against the real HF weights: ragged univariate + multivariate inputs with multiple cutoffs produce correctly shaped (n_cutoffs, H, C, Q) outputs, finite values, monotone quantiles, and correct prediction_length override + empty-input handling.
• benchopt test . -k T0: passes the same tests as the existing chronos2 solver (test_solver_class, test_solver_install_api); the remaining failures are the missing-test-dataset issue on main that ENH Fix running the test suite #44 fixes, and reproduce identically for chronos2.
• ruff check clean.

🤖 Generated with Claude Code

[felixdivo]

Awesome addition! It probably makes sense to wait a tiny bit to then build this directly on the simpler implementation provided by #42. This is less relevant for T0 because it does not expose forecasts, but it would still enable anomaly detection via residual error tracking.