REFACTOR: unify error/blocked response scoring across scorers

[romanlutz]

Description

Blocked (content-filtered) responses were being handled in three independent places that all converged on 0.0 / False by accident: the TrueFalseScorer no-pieces fallback, the FloatScaleScoreAggregator empty-list fallback, and TAP's error_score_map. Two scorers also had broken edge cases — SelfAskCategoryScorer would send error content to the LLM (likely raising InvalidJsonException or hallucinating a category), and ConversationScorer would crash with ValueError on any blocked turn.

This PR pushes blocked handling into the scorer base classes so the "correct" default behavior is what every scorer gets for free:

• TrueFalseScorer defaults to False on blocked input (existing fallback, now load-bearing).
• FloatScaleScorer defaults to 0.0 on blocked input (existing aggregator fallback, now load-bearing and documented).
• SelfAskCategoryScorer's validator now rejects error pieces so it gets the same False default instead of asking the LLM about garbage.
• ConversationScorer no longer rejects error pieces wholesale, so blocked turns mid-conversation are scored normally.
• TAP's error_score_map is removed — it was added after the last release, was redundant with the new defaults, and only confused things further. SelfAskRefusalScorer still inverts True/False semantics on blocked content (refusal detected → True); that's intentional and the only "weird" default a user must still wrap in TrueFalseInverterScorer when using it as an objective scorer.

Tests and Documentation

• New tests for the unified defaults in test_float_scale_threshold_scorer.py, test_scorer.py, test_self_ask_category.py, test_conversation_history_scorer.py, and the existing TAP suite.
• TAP notebook (doc/code/executor/attack/tap_attack.{py,ipynb}) re-executed end-to-end against the real Azure OpenAI GPT-4o strict-filter target and OpenAI image target. The new "The request was blocked by the target; returning 0.0." rationale now appears in the tree output, confirming the refactor fires in a live run.
• Class docstrings on FloatScaleScorer and TrueFalseScorer updated and converted from RST to MyST markdown so they render correctly in the jupyter-book docs site.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[behnam-o]

I like the unification by introducing the idea of fallback score, and handling in the superclass. I just think the fallback score, by definition, should always have a value, and it can't be optional, or None. But I might be missing nuances, so not blocking.

[romanlutz]

Self-correction: reverted the TrueFalseInverterScorer change from commit 5d0f5e7. On further review, the original passthrough-and-invert behavior is correct. When the inner scorer's fallback fires False ("attack did not succeed"), the inverter correctly inverts that to True under the inverter's flipped semantic ("attack failed" or "no compliance detected"), which is the right answer for blocked messages. Forcing the inverter's own fallback instead would have flipped this to a misleading False.

The ConversationScorer fix from the same commit is unaffected and still in place.