bench: LM Studio / Ollama latency+tok/s comparator as a native Tool

[MauricioPerera]

Repo: https://github.com/MauricioPerera/bench

Third plugin in this batch (after kdd-gates and gh-discussions): bench_run compares latency and tokens/sec on the identical prompt across LM Studio (local, OpenAI-compatible /v1/chat/completions) and Ollama (local or :cloud models, native /api/chat). Pure fetch(), no subprocess involved.

Two real bugs found while verifying it live, might save someone else the trace:

• dsh-tools validates a tool's return value as "lossless JSON" and rejects undefined as a property value — JSON.stringify silently drops those, but this validator throws "value is not lossless JSON" instead. Use null for a field you don't have, not undefined.
• Ollama's :cloud models (tested with gpt-oss:20b-cloud) don't return the eval_duration/prompt_eval_duration breakdown that local models do — only total_duration. Worth a fallback if you're computing tok/s from it.

Runs targets sequentially (parallel local calls would contend for the same GPU/CPU and skew the timing). MIT licensed, tagged dsh-plugin.