Flatten Claude trace spans for user turns

[adagradschool]

Summary

Update the Claude Code plugin for the flattened trace UI:

• remove the synthetic Turn span
• emit tool and LLM spans directly under the session span
• label the first span after human input as LLM Call (User)
• make user prompt text render more cleanly in flattened spans

Why

In the flattened view, the turn wrapper adds noise. This makes each user prompt show up as its own top-level call with the first resulting Claude response or tool call, without backend changes.

Testing

• python3 -m py_compile plugins/pl-trace-claude-code/hooks/parse_stop_transcript.py
• make test