feat: Add dedicated GitHub Actions Performance Benchmarking workflow

[solderzzc]

Overview

This PR introduces a dedicated, standalone GitHub Actions workflow (.github/workflows/benchmark.yml) designed to automatically benchmark Apple MLX generation speeds directly on the CI infrastructure (macos-15 runner).

Since Cloud macOS runners typically possess exactly 7 GB of RAM, blindly running the standard benchmark script matrix risks immediate Out-Of-Memory (OOM) SIGKILL boundaries. This safe workflow explicitly forces constraints and exposes a configurable matrix to test safe parameter bounds directly from the UI.

Features

• UI Triggered (workflow_dispatch): Fully configurable directly through the GitHub Actions web UI. Does not run implicitly on pushes.
• Configurable Models: Supports swapping the model_id input immediately prior to run. Defaults to the un-gated mlx-community/gemma-4-e4b-it-4bit which safely drops into memory constraints.
• Context Boundaries: Tests generation strictly against safe, small context lengths ("512,1024,4096") by default.
• SSD streaming isolation: Toggles the --ssd-only parameter to explicitly bypass physical limits and test extreme combinations.

Outputs

• Downloads and manages dependencies (mlx, psutil, requests) inside the GH Action.
• Bundles the final generated github-action-benchmark.md table securely as an exported execution artifact on each run.