FEAT: add MessageToolCallScorer for tool calls recorded in stored messages

[WatchTree-19]

Description

Adds MessageToolCallScorer, the message-side counterpart to OtelToolCallScorer. It scores the same ToolsCalled condition from the function-call pieces PyRIT already stores, so it works without a trace pipeline.

A tool counts when a model-authored function_call piece is paired by call ID with a later function_call_output piece in the scored message's conversation, up to and including that message, which is the same scope resolve_message_trace_scope uses. Both the Responses (call_id, name) and Chat Completions (id, function.name) formats are read. Requested calls with no output do not count, calls in simulated_assistant history are ignored, and PyRIT's tolerant-mode dispatch outputs (function_not_found, missing_function_name, malformed_arguments) do not count, since the function never ran. An error returned by the tool itself still counts, in line with ToolsCalled counting an attempt regardless of success.

Stored messages are partial evidence: hosted tools and several response section types are not persisted, and targets that run tools themselves store no outputs. The scorer therefore returns true when every required tool ran and undetermined otherwise, never false, and it composes with OtelToolCallScorer under TrueFalseScoreAggregator.OR. It does not create observations, so score_observation_async is not supported for it.

Tests and Documentation

• tests/unit/score/test_message_tool_call_scorer.py (30 tests): both call formats, unpaired requests, dispatch errors versus tool-reported errors, call ID pairing, output ordering, later turns and other conversations, role and data type filtering, multiple required tools, an OR composite with OtelToolCallScorer, and an end-to-end PromptSendingAttack through a mocked OpenAIResponseTarget tool loop.
• Each check in the scorer was mutation-tested: all 11 deliberate breaks fail at least one test.
• tests/unit/score, tests/unit/docs and tests/unit/registry pass; ruff and ty are clean.
• Added a markdown section to doc/code/scoring/5_tool_call_scorer.py and synced the .ipynb with JupyText. No code cells changed, so no re-execution was needed.

[WatchTree-19]

Rebased onto main after #2852 and #2832. The scorer now reads calls and outputs through the shared FunctionCallContent and FunctionOutputContent models instead of its own JSON parsing, uses the async memory API, and ignores simulated_tool outputs as well as simulated_assistant calls, with a new test for that case. 31 tests in test_message_tool_call_scorer.py pass, and tests/unit/score is green (2926 passed).