[ENH] Allow `TorchExperiment` metric to set "higher is better"

[SimonBlanke]

The TorchExperiment assumes all metrics should be minimized (lower is better) by default. This can be seen here in this line. However, many metrics should be maximized (accuracy, F1, AUC), requiring users to manually negate them or wrap them in custom logic. The integration doesn't auto-detect metric direction from common naming patterns.

Sklearn's OptCV uses _coerce_to_scorer_and_sign (from _skl_metrics.py) to automatically detect metric direction. Maybe we can write something similar for the TorchExperiment.

[amitsubhashchejara]

Okay I will check this.

[sriharan0804]

Hi @SimonBlanke, I’d like to work on this.

My plan is to add a backward-compatible way for TorchExperiment metrics to specify direction, probably with a higher_is_better option while keeping the current minimize behavior as default unless auto-detection is preferred. I’ll also check the existing sklearn OptCV _coerce_to_scorer_and_sign logic and add tests for metrics like loss, accuracy, F1, and AUC.
Please let me know if this direction sounds good.