Benchmark Script for local models

[LeeSinLiang]

Benchmark Script for local models

Summary
Add scripts/benchmark.py to compare local HF models on a fixed diff corpus.

Scope

• Inputs: a small corpus of representative diffs (JSONL).
• Outputs: CSV with model, tokens_in, tokens_out, latency_ms, provider_device.
• Optional: save a plot to docs/benchmarks/.

Tasks

• Implement benchmark harness with warmup and N runs.
• Detect device (CUDA/MPS/CPU) and memory optimization flags.
• Add sample corpus + instructions.
• Document results format in README.

Acceptance criteria

• Running the script produces a CSV and a short summary line per model.
• Works without cloud keys (local models only by default).