Add submetrics for more metrics

[gabegma]

Summary

Adds sub-metric breakdowns to existing metrics so results can be analyzed at a finer grain (per error type, per failure mode, per dimension, per entity type) without having to add new top-level metrics. Also threads a higher_is_better signal through the metric base class, aggregator, and analysis app so lower-is-better metrics are rendered correctly.

Framework changes

• BaseMetric.higher_is_better class attribute. Default True; override on the class (e.g., latency metrics) rather than per record.
• New helpers in eva.metrics.utils:
  • make_rate_sub_metric(...) — rate-style sub-metric where score == normalized_score == numerator / denominator.
  • build_binary_flag_sub_metrics(...) — one sub-metric per dimension/corruption type with score = 1.0 when the judge flagged it (aggregated mean reads as an issue-occurrence rate).
  • direction_for_sub_metric(...) — derives direction from the key suffix: _rate → lower-is-better, _accuracy → higher-is-better, otherwise inherit the parent.
• PerTurnConversationJudgeMetric.build_sub_metrics(context, per_turn_ratings, per_turn_extra) hook for subclasses to surface per-turn breakdowns.
• MetricsRunner now writes higher_is_better on every metric and sub-metric entry in the run-level aggregates, reading the parent direction from the registry.

Per-metric sub-metrics added

• stt_wer: substitution_rate, deletion_rate, insertion_rate (component / reference-word count).
• tool_call_validity: num_tool_calls (count, non-normalized) + one *_rate per error type in CALL_ERROR_TYPES.
• conciseness_judge: one <failure_mode>_rate per failure mode, over rated turns.
• conversation_progression_judge: one <dimension>_rate per dimension (binary flagged).
• faithfulness_judge: one <dimension>_rate per dimension (binary flagged).
• transcription_accuracy_key_entities: one <entity_type>_accuracy per entity type seen in the run.
• user_behavioral_fidelity: one <corruption_type>_rate per corruption type (binary detected).

Analysis app

• Prefixes lower-is-better labels with ↓ (parent direction from the registry; sub-metric direction from the key suffix).
• Supports non-normalized sub-metrics like tool_call_validity__num_tool_calls (count axis, no hover suffix).
• Bar ordering, query-param-driven state, and error surfacing in the analysis view.

Test plan

• pytest tests/unit/metrics — unit tests added for each metric's sub-metric output (conciseness, conversation progression, faithfulness, STT WER, tool call validity, transcription accuracy key entities, user behavioral fidelity) and for the runner's sub-metric aggregation with higher_is_better.
• Spot-check the analysis app against a recent run — confirm ↓ prefix on lower-is-better metrics and that tool_call_validity__num_tool_calls renders on the count axis.