Align ToolCallAccuracy output schema to reason and harden skipped/properties handling

[Copilot]

This update aligns the Tool Call Accuracy evaluator with the PR-expected response contract by renaming reasoning to reason across evaluator output, prompt schema, and scoped tests. It also corrects not-applicable payload shape (properties: null) and prevents runtime issues when LLM returns properties: null.

• Contract updates in evaluator output

  • _return_not_applicable_result now emits:
    • tool_call_accuracy_reason (was tool_call_accuracy_reasoning)
    • tool_call_accuracy_properties: None (was {})
  • ToolCallAccuracyEvaluator now reads llm_output["reason"] and emits ..._reason.

• Null-safe properties merge

  • Updated properties extraction to handle nullable payloads safely:

  llm_properties = llm_output.get("properties", {}) or {}

• Prompt/schema alignment

  • tool_call_accuracy.prompty now uses reason (not reasoning) in skipped example and output-key instructions.
  • Output key order was updated so reason appears before score, matching requested schema ordering.

• Scoped test updates for this PR surface

  • Updated relevant Tool Call Accuracy unit tests from reasoning/_reasoning to reason/_reason.
  • Kept changes limited to tests touched by this PR scope.