fix: failed_count is always zero in result analysis

[WatchTree-19]

problem

in compute_confusion (delphi/log/result_analysis.py), failed_count is computed as len(df_valid) - total, but total is itself len(df_valid), so failed_count is always 0. as a result the "Average fraction of failed examples" that log_results prints is always 0.0, even when the scorer failed to produce a (parseable) prediction for many examples (those rows are dropped by the prediction.notna() filter earlier in the function).

fix

-        failed_count=len(df_valid) - total,
+        failed_count=len(df) - total,

len(df) - total is the number of rows removed by the .notna() filter, i.e. the examples that failed.

test

added tests/test_log/test_result_analysis.py (pandas-only): builds a frame with 2 NaN predictions (what classifier.py stores when generation/parse fails) and asserts failed_count == 2 and fraction 0.5. fails on current main (assert 0 == 2), passes with the fix. ruff clean.

found by reading the code.