feat(iterate2): support per-metric direction (minimize/maximize)

[romeokienzler]

Summary

Adds support for specifying whether each metric should be minimized or maximized in iterate2. Previously all metrics were unconditionally maximized.

Changes

terratorch_iterate/iterate2/_iterate2.py

• _default_direction(metric_name) — new heuristic helper: metrics whose name contains loss, error, err, mse, mae, or rmse default to minimize; all others default to maximize.
• load_metrics() — now returns (names, directions) tuple. The metrics: YAML key now accepts two forms:
  • Simple string — direction inferred by name heuristic
  • Extended dict {name: ..., direction: minimize|maximize} — explicit per-metric control
  • A mix of both is allowed; invalid direction values raise ValueError.
• main() — unpacks the tuple and passes the real directions to optuna.create_study() instead of the hardcoded ["maximize"] * len(metrics).

docs/iterate2.md

• Added "Metric directions (minimize / maximize)" subsection under "Metric extraction" documenting the heuristic, both YAML forms, and a mixed-form example.
• Removed the stale "All objectives are maximised" statement.

YAML usage

Simple form (direction inferred from name):

metrics:
  - val_loss      # → minimize  (contains "loss")
  - accuracy      # → maximize

Extended form (explicit per-metric):

metrics:
  - name: val_loss
    direction: minimize
  - name: accuracy
    direction: maximize

Mixed form is also supported.