Add Python inference benchmark CLI

[RitwijParmar]

Adds a cactus benchmark command for repeatable local inference measurements from the Python CLI.

Why this felt useful: the Python API already returns TTFT, prefill throughput, decode throughput, RAM, and token counts, but there was not a small built-in way to run the same prompt profiles repeatedly and save artifacts for comparing runtime, quantization, or long-context changes.

What changed:

• adds built-in benchmark profiles for short chat, long prefill, and JSON-shaped output
• supports custom profile files so a PR or device run can publish exact prompt shapes
• writes per-run JSONL plus grouped summary JSON with mean, p50, p95, min, and max
• resets the KV cache between runs by default, with --keep-cache when cache reuse is intentional
• documents the workflow in the Python README

I used Codex to help implement and test this, and reviewed the code locally before opening the PR.

Tests:

• /Users/ritwij/Documents/oss-work/cactus/.venv-codex/bin/python -m pytest python/tests/test_cli_benchmark.py python/tests/test_cli_run.py -q
• /Users/ritwij/Documents/oss-work/cactus/.venv-codex/bin/python -m pytest python/tests/test_bindings.py::TestCliParser python/tests/test_cli_benchmark.py -q

Both passed locally before I removed the temporary test venv.

[RitwijParmar]

I pushed a follow-up that turns this into more of a regression harness rather than just a runner.

New pieces:

• synthetic context sweep profiles via --sweep-token-counts
• environment metadata in summaries so runs are easier to compare later
• Markdown reports for benchmark artifacts
• --compare base.json candidate.json with thresholded regression checks for TTFT, total latency, prefill TPS, and decode TPS
• compare mode avoids importing the native binding, so it can run anywhere the summary artifacts are available

I also added tests around the sweep generation, comparison logic, Markdown output, and parser paths.

[RitwijParmar]

Rebased this on current main and resolved the CLI conflicts. The benchmark command now sits next to the newer run/list/test CLI changes instead of replacing them. I also checked the touched Python files with Python 3.11 compile. Local pytest is not available in this environment, so I could not run the focused pytest here.