fix(claude-agent-sdk): Fix subagents span nesting and input

[Luca Forstner (lforst)]

Example stream:

Root user:
  "Solve this using a math sub-agent"

Root assistant:
  calls Agent tool with prompt:
  "Add 15 and 27"

Sub-agent user message, parent_tool_use_id=agent-1:
  "Add 15 and 27"        <-- this is the sub-agent prompt

Sub-agent assistant:
  calls calculator tool

Sub-agent user message, parent_tool_use_id=agent-1:
  tool_result: "42"

Prev (bad):

root trace
├─ root LLM
│  └─ Agent(prompt: "Add 15 and 27")
└─ sub-agent
   ├─ LLM input: "Add 15 and 27"
   └─ LLM input: "42"                 <-- wrong: tool result treated like prompt

Now (good):

root trace
├─ root LLM
│  └─ Agent(prompt: "Add 15 and 27")
└─ sub-agent
   ├─ LLM turn 1
   │  ├─ input:
   │  │  └─ user: "Add 15 and 27"
   │  └─ output:
   │     └─ assistant tool_use: calculator
   └─ LLM turn 2
      ├─ input:
      │  ├─ user: "Add 15 and 27"
      │  ├─ assistant tool_use: calculator
      │  └─ user tool_result: "42"
      └─ output:
         └─ assistant: "The answer is 42"

Fixes #2103

[Abhijeet Prasad (AbhiPrasad)]

seems like preferTaskSiblingParent is always true now? Do we even need an option?