fix(claude_agent_sdk): preserve tool span output across back-to-back tool_use messages

[Abhijeet Prasad (AbhiPrasad)]

resolves #318

AI Summary

In claude-agent-sdk 0.1.61 a delegated subagent can emit multiple tool_use AssistantMessages (e.g. Bash then Read) inside one LLM turn before any tool_result arrives. ContextTracker._handle_assistant was unconditionally calling cleanup_context whenever the context already had an llm_span, which closed the in-flight Bash tool span with no output. When the matching tool_result finally arrived two events later, the span was already gone and its output was dropped (issue #318).

Only run cleanup_context when ctx.next_llm_start is set, which is the real LLM-turn boundary (set in _handle_user when a tool_result arrives). Back-to- back tool_use messages now merge into the existing LLM span without closing sibling tool spans.

• Red/green regression: new ContextTracker unit test simulating the 0.1.61 event stream (Bash tool_use -> Read tool_use -> Read tool_result -> Bash tool_result, all under one subagent parent).
• Re-records cassettes under claude-agent-sdk 0.1.61 and bumps the matrix latest pin to 0.1.61. test_interleaved_subagent_tool_spans_preserve_output and test_interleaved_subagent_tool_spans_parent_to_correct_llm now pass.
• Full test_claude_agent_sdk(latest) and test_claude_agent_sdk(0.1.10) sessions green.