feat: sobol_indices() -- expose total-order Sobol indices (#120)

[jc-macdonald]

Summary

`screen(method="sobol")` only ever returned S1 (first-order), silently discarding ST (total-order) even though the same underlying `SALib.analyze.sobol.analyze()` call already computes it for free. `ST - S1` is the standard way to detect interaction effects -- a factor with small S1 but large ST is interacting with other factors rather than acting independently -- and `screen()` couldn't surface that at all.

Design

• New `sobol_indices(run_fn, factors, *, n_samples=100, seed=42) -> dict[str, tuple[NDArray, NDArray]]`, returning `{observable: (S1, ST)}`. `screen()` itself is untouched, so this is fully backward compatible for existing callers.
• Shares the Saltelli sampling/evaluation loop with `_screen_sobol` via a new `_sobol_sample_and_evaluate()` helper rather than duplicating it -- both draw from and evaluate the exact same design given the same inputs, confirmed by a test asserting `sobol_indices`'s S1 matches `screen()`'s S1 exactly.

Test plan

• `just ci` green (345 passed, `design.py` 100% coverage)
• `sobol_indices`'s S1 output matches `screen(method="sobol")`'s S1 exactly (same seed/sampling)
• `ST >= S1` holds (within MC estimation tolerance)
• From-scratch interaction-detection test: a pure product of two zero-mean-centered factors (`E[Y|A] = A*E[B] = 0` and vice versa, so S1~0 for both) shows `ST > 0` for both, confirming ST actually catches a real interaction S1 structurally can't see -- not just a shape/type check
• Rejects a factor list with no continuous factors, matching `screen()`'s existing behavior

Closes #120