feat: dedicated LLM/AI agent health view in TUI and CLI

[planetf1]

otelite already collects rich GenAI semantic convention attributes (gen_ai.request.model, cache_read_tokens, cache_creation_tokens, ttft_ms, success, error, streaming state). The otelite usage subcommand surfaces token stats, but there is no view oriented towards diagnosing AI agent session failures.

When debugging a Claude Code session that was producing repeated Internal server error responses via a LiteLLM → Bedrock proxy, the diagnostic questions were:

• Which sessions/interactions failed, and with what model?
• What was the token count and body size at the point of failure?
• Was the error a streaming stall or an immediate rejection?
• What request IDs should I give to my proxy/cloud provider for escalation?

These questions required manually stitching together traces list, traces show, logs search, and logs show across 4–5 commands.

Requested behaviour:

A otelite llm subcommand (or a TUI pane) oriented around AI agent health:

• Table of recent LLM requests: timestamp, model, tokens (input/cached/output), TTFT, success/error, request ID
• Highlight failing requests; show error message inline
• Drill-down: select a request → show associated spans + request/response body logs
• Filter by model, session, status

This maps well to the existing claude_code.* span taxonomy Claude Code emits, but would be useful for any GenAI workload emitting OTel semantic conventions.