fix: resolve TypeError during leaderboard sort on AutoML reload (fixes #681)

[manoprasad2006]

Summary

Fixes #681 — AutoML reload crashes with TypeError: '<' not supported between instances of 'float' and 'str' during model stacking or
leaderboard sorting.

Root Cause

1. ModelFramework.save() serialized final_loss to framework.json
   as a string via str(desc["final_loss"]).
2. ModelFramework.load() read it back as a raw string (e.g. "0.1234").
3. When new models were trained alongside reloaded ones, the leaderboard
   mixed float (new models) and str (reloaded models) in metric_value.
4. ldb.sort_values(by="metric_value") crashed because Python 3 does not
   support < comparisons between float and str.

Changes

• supervised/model_framework.py
  • Removed str() cast from final_loss in ModelFramework.save()
    so it is written as a numeric float
  • Added float() conversion in ModelFramework.load() with a
    try-except for backward compatibility with older saved directories
• supervised/ensemble.py
  • Added float() conversion in Ensemble.load() when reading
    best_loss from ensemble.json for extra robustness

Testing

Verified by reloading a saved AutoML directory and continuing training —
leaderboard sorts correctly with no TypeError.

Closes #681