Add a built-in profiling callback for per-invocation timing analysis

[acere]

Problem

When investigating where time is spent during an LLMeter run, users currently need to write custom ad-hoc callbacks (as seen in dev notebooks) to capture per-request phase timings. There is no built-in way to:

1. Break down request time into server-side phases — specifically separating prefill (TTFT) from generation (decode), and distinguishing server time from client-side runner overhead.

2. Capture per-invocation metadata alongside timing — cache hit status, reasoning token counts, retry counts, and throughput metrics — in a structured format suitable for analysis.

3. Persist profiling data to disk for post-hoc analysis without needing the live callback or endpoint.

4. Visualize profiling results with ready-made plots (phase breakdown, request timeline/concurrency, throughput over time, TPOT distribution, etc.).

Expected Behavior

A built-in ProfileCallback that:

• Hooks into the existing callback lifecycle (before_invoke / after_invoke / after_run)
• Captures detailed per-invocation profiles (timing phases, token counts, cache info, retries)
• Produces an aggregated report with time accounting and percentile statistics
• Saves structured data (JSON + JSONL) alongside run results
• Contributes summary stats to result.stats for programmatic access
• Pairs with plotting utilities for common profiling visualizations

Use Case

Understanding where wall-clock time goes in a benchmark run — especially distinguishing server-side prefill and decode from client-side overhead — and identifying patterns like cache hit impact on TTFT, throughput degradation over time, or concurrency effects.