[FEAT]: Manually scoring messages and attacks via GUI

[jbolor21]

Description

Separates automated and human scoring on attack results and keeps manual scoring focused on objective true/false verdicts.

• Adds automated_score and human_score to attack results.
• Makes last_score a read-only effective-score property with human-score precedence.
• Migrates existing last_score_id values to automated_score_id and adds human_score_id.
• Updates attack executors to write objective scorer output to automated_score.
• Adds boolean manual scoring with an optional rationale and authenticated user identifier metadata.
• Adds an explicit, enabled-by-default Update attack score and outcome option.
  • When selected, only human_score, outcome, and outcome_reason are updated.
  • When not selected, the score is retained only in message score history.
• Validates that the scored message belongs to the supplied attack before objective checks or updates.
• Supports manual scoring for messages in the attack's forked conversations.
• Extends attack PATCH support to update the shared objective and its hash.

Tests

• Added migration and score-precedence coverage.
• Added backend coverage for manual scoring, ownership validation, and forked conversations.
• Added frontend coverage for boolean scoring and attack update behavior.
• Ran impacted Python tests, focused frontend Jest suites, Ruff, ESLint, TypeScript type-checking, and git diff --check.