gooddata-eval: join forecasting turns to the item trace scope #1941

CodeRabbit finding on #1798, not addressed before merge (non-blocking, tracked here instead): #1798 (comment)

packages/gooddata-eval/src/gooddata_eval/core/agentic/forecasting.py:510-513 opens its trace with open_trace_window directly instead of open_item_trace/item_scope. When a run needs a clarification turn, forecasting can score across turns that aren't joined under one item trace root, so Langfuse exports a split/wrong trace for that item.

Fix direction (from the review): switch to open_item_trace + item_scope, pass run.conversation_id through to ctx.observe, mirroring the pattern kda_skill.py already uses.