feat: add ProfileCallback for per-invocation profiling

[acere]

Closes #90

Summary

Adds a built-in ProfileCallback that captures per-invocation timing data during a Runner run and produces a structured profiling report.

Changes

New: llmeter/callbacks/profiling.py

• InvocationProfile dataclass — per-request record with timing phases, token counts, cache info, retries
• ProfileReport dataclass — aggregated summary with time accounting, cache hit rate, retry stats, percentile statistics
• ProfileCallback — lifecycle hooks that collect data and generate the report in after_run
• Saves profile_report.json + profile_invocations.jsonl alongside run results

New: llmeter/plotting/profiling.py

• plot_phase_breakdown — stacked bar (TTFT vs generation per request)
• plot_request_timeline — Gantt-style concurrency view
• plot_throughput_over_time — tok/s scatter with rolling average
• plot_time_accounting — pie chart (server prefill vs decode vs client overhead)
• plot_tpot_distribution — TPOT histogram
• plot_ttft_vs_input_tokens — scatter with cache hits highlighted
• plot_profile_summary — 2×2 multi-panel dashboard

Supporting changes

• llmeter/callbacks/__init__.py — exports ProfileCallback
• llmeter/plotting/__init__.py — exports profiling plot functions
• mkdocs.yml — nav entries for new API reference pages
• docs/reference/callbacks/profiling.md + docs/reference/plotting/profiling.md
• examples/Profiling.ipynb — full walkthrough notebook

Tests

• tests/unit/callbacks/test_profiling.py — 29 tests covering serialization, lifecycle, report computation, stats contribution, file saving, and edge cases

Terminology

Clear server/client separation in naming:

• TTFT = server prefill + network round-trip
• Generation time = server-side decode (TTLT − TTFT)
• Runner overhead = client-side (queue, token counting, callbacks)

Testing

uv run pytest tests/unit/callbacks/test_profiling.py -v  # 29 passed
uv run ruff check llmeter/callbacks/profiling.py llmeter/plotting/profiling.py

[athewsey]

(Note, seems like this PR depends on plotting.defaults introduced in #89)

Just seen a weird intermittent error while testing the example notebook here... Requests failing with:

Endpoint invocation failed: 'dict' object has no attribute 'id'
Traceback (most recent call last):
  File "{...}/llmeter/llmeter/endpoints/base.py", line 259, in wrapper
    self.process_raw_response(raw_response, start_t, response)
  File "{...}/llmeter/llmeter/endpoints/openai_response.py", line 281, in process_raw_response
    response.id = event.response.id
                  ^^^^^^^^^^^^^^^^^
AttributeError: 'dict' object has no attribute 'id'

The same happens with google.gemma-4-31b on the same new /openai/v1 endpoint, but interestingly not with openai.gpt-oss-120b on the old /v1 endpoint. Claude couldn't figure yet what path through the OpenAI Python SDK could possibly be allowing this object to pass through un-parsed.

Will come back to it later