v0.6.1: fix dataset configs missing from wheel

The *.json entry in .gitignore caused hatchling to exclude
examples/dataset_configs/*.json from the built wheel, even though
the files were force-tracked in git. Added a negation pattern to
exempt the bundled dataset configs.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

v0.6.0: non-streaming default, -n completions, vertical plot layout

- Default to non-streaming for higher throughput
- Add -n/--num-completions for RL rollout generation
- Add --stream flag to opt into SSE streaming
- Plot: vertical 2-panel layout (throughput + latency) for non-streaming
- Plot: cleaner titles

Bump version to 0.5.4 — fix bundled dataset configs missing from wheel

Bump version to 0.5.3

Bump version to 0.5.2

Bump version to 0.5.1

Update README for v0.5.0: CLI commands, --results-dir, pip install

Fix steady-state metric: 50ms buckets and better filtering

- Reduce bucket size from 1s to 50ms for accurate steady-state measurement
- Add target_concurrency param to avoid bucket-boundary peak inflation
- Remove queue wait time from TTFT in sustained mode
- Adjust default prompts per sweep to 8*concurrency

New example benchmark command

v0.1.0: Initial release

Includes:
- completion_simple_benchmark.py: Basic async throughput benchmark
- completion_advanced_benchmark.py: Streaming benchmark with TTFT/TPOT/ITL metrics
- embedding_benchmark.py: Concurrent embedding throughput benchmark
- Scenario-based sampling (Normal, Deterministic, Uniform distributions)
- LoRA benchmarking support with multiple distribution strategies
- Multi-process execution with deterministic seeding
- Plotting and visualization tools