Auto-generate release notes

[rviscomi]

Fixes an open TODO to generate release notes, using the Gemini API in the publish workflow.

For testing, you can generate a dry run of the release notes with this command:

pnpm --filter serving run publish-skills --dry-run

Or to test/backfill previous releases, pass the tag ID to the script directly:

node --env-file=.env --experimental-strip-types serving/skills-cli/generate-release-notes.ts v0.0.184

Example of release notes generated for today's 0.0.185 release:

# Release Notes: `v0.0.185`

### 📖 Guidance & Web Platform Updates

#### 🔄 Guide Updates & Platform Evolution
* The **Translator API** guide has been updated to mandate usage of the global `Translator` interface over the deprecated `window.ai.translator` namespace and clarifies strict user gesture requirements and progress monitoring during model downloads.

### 📊 Benchmark Evaluations
| Agent + Model | Tasks / Assertions | Unguided → Guided Pass Rate | Uplift |
| :--- | :---: | :---: | :---: |
| **claude_code** (claude-opus-5) | 130 / 1033 | 58% → **92%** | **+34pp** |
| **antigravity** (Gemini 3.7 Flash Preview) | 130 / 1112 | 64% → **90%** | **+26pp** |
| **codex_cli** (gpt-5.6-sol) | 130 / 1112 | 60% → **83%** | **+23pp** |

---
**Full Changelog**: https://github.com/GoogleChrome/modern-web-guidance/compare/v0.0.184...v0.0.185

Another example using a recent release that added new use cases:

# Release Notes: `v0.0.178`

### 📖 Guidance & Web Platform Updates

#### 🆕 New Guides
* **persistent-toast-notifications**: Provides best practices for implementing accessible, persistent feedback alerts using modern web patterns.
* **highlight-text-ranges**: Covers the use of modern CSS to highlight arbitrary text ranges for features like search results, spelling errors, or collaborative editing cursors.

#### 🔄 Guide Updates & Platform Evolution
* **navigation-drawer**: Relocated to the `user-experience` category to better align with modern UX architectural patterns.
* **Lazy-loading images and iframes**: Expanded platform coverage to include native lazy-loading features for optimized resource management.

### 📊 Benchmark Evaluations
| Agent + Model | Tasks / Assertions | Unguided → Guided Pass Rate | Uplift |
| :--- | :---: | :---: | :---: |
| **claude_code** (claude-opus-5) | 130 / 1033 | 58% → **92%** | **+34pp** |
| **antigravity** (Gemini 3.7 Flash Preview) | 130 / 1112 | 64% → **90%** | **+26pp** |
| **codex_cli** (gpt-5.6-sol) | 130 / 1112 | 60% → **83%** | **+23pp** |

---
**Full Changelog**: https://github.com/GoogleChrome/modern-web-guidance/compare/v0.0.177...v0.0.178