MAINT: Detangle scorer LLM round-trip — ResponseHandler + internal helper (1/3)

[romanlutz]

This is step 1 of 3 in a scorer-architecture refactor — "Option 5: Scorer-owned composition". Full design notes: design gist.

Follow-up to #1867, which bolted a response_parser hook onto SelfAskTrueFalseScorer for LlamaGuard and surfaced that the base Scorer class conflates concerns that should be separate (evaluation mechanism, system-prompt construction, and response parsing all live on the base, so every non-LLM scorer inherits machinery it can never use).

Scope of this PR — internal refactor, no public API change

No behavior change and no constructor changes. This purely rehomes internals so the base Scorer stops carrying LLM-only machinery.

• ResponseHandler abstraction (pyrit/score/response_handler.py) — a ResponseHandler ABC plus a default JsonSchemaResponseHandler that owns turning the target's raw text into the parsed UnvalidatedScore shape (score_value, rationale/description, metadata, category). JsonSchemaResponseHandler reproduces today's parsing exactly — json.loads(remove_markdown_json(...)) plus the same key lookups, category/metadata normalization, and InvalidJsonException/ValueError behavior.
• Stateless round-trip helper (pyrit/score/llm_scoring.py) — a module-internal _run_llm_scoring_async(*, chat_target, system_prompt, response_handler, value, ...) that performs the LLM round-trip previously living in Scorer._score_value_with_llm_async: sets the system prompt on the target, sends the prompt, applies the existing @pyrit_json_retry behavior, and delegates parsing to the handler.
• Base Scorer slimmed — the retry decorator, JSON parsing, and system-prompt handling move off the base. Scorer._score_value_with_llm_async is now a thin internal forwarder to the helper, kept only so the not-yet-migrated scorers keep working. It is deleted in step 3.
• FloatScaleScorer override removed — the old override only added a float() post-check on the parsed value; that check is folded into the helper via a _score_value_is_numeric flag, preserving behavior.
• SelfAskTrueFalseScorer wired to call the helper directly. Its public constructor and behavior are unchanged.

The 3-step plan (per the gist)

1. This PR — introduce the internal round-trip helper + ResponseHandler (JSON-schema default), rehome the _score_value_with_llm_async logic off the base Scorer, and wire SelfAskTrueFalseScorer. No public-API change.
2. PR B — expose the composition on SelfAskTrueFalseScorer.__init__ (chat_target + system_prompt + response_handler) plus a TrueFalseQuestion value type; add the CallableResponseHandler escape hatch and pilot LlamaGuard. from_question_yaml becomes a deprecation shim.
3. PR C — roll the pattern out to the remaining LLM scorers, one at a time, and delete the transitional base forwarder.

Composites/wrappers (TrueFalseCompositeScorer, Audio/VideoTrueFalseScorer, ConversationScorer, …) and all non-LLM scorers are unchanged.

Testing

• uv run pytest tests/unit/score → 1232 passed, 16 skipped.
• ruff, ty, and pre-commit all clean.