FIX: Move AttackStrategy import to TYPE_CHECKING to break circular import

PR #1881 (Pydantic identifiers refactor) re-merged with this PR's
compute_inner_attack_eval_hash addition and hoisted
rom pyrit.executor.attack.core.attack_strategy import AttackStrategy
to module level. That import triggers a cycle through
`pyrit.executor.attack` -> `pyrit.message_normalizer` ->
`pyrit.common.data_url_converter` -> `pyrit.models`, which fails
because `pyrit.models` is still being initialised at that point
(DataTypeSerializer is not yet defined).

`from __future__ import annotations` is already enabled, so the
type annotation `attack: AttackStrategy` works as a string. Moving
the import inside `if TYPE_CHECKING:` restores the original lazy
boundary and unblocks CI.

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

Author: hannahwestra25

pyrit/models/identifiers/evaluation_identifier.py

@@ -23,13 +23,15 @@
 from __future__ import annotations
 
 from abc import ABC
-from typing import Any, ClassVar, Optional
+from typing import TYPE_CHECKING, Any, ClassVar, Optional
 
 from pydantic import BaseModel, ConfigDict, Field
 
-from pyrit.executor.attack.core.attack_strategy import AttackStrategy
 from pyrit.models.identifiers.component_identifier import ComponentIdentifier, config_hash
 
+if TYPE_CHECKING:
+    from pyrit.executor.attack.core.attack_strategy import AttackStrategy
+
 # Behavioral params that define model output quality for scoring.
 TARGET_EVAL_PARAMS: frozenset[str] = frozenset({"underlying_model_name", "temperature", "top_p"})
 TARGET_EVAL_PARAM_FALLBACKS: dict[str, str] = {"underlying_model_name": "model_name"}