docs(skill): refresh on-device-ai skill for v0.9.0 (#1186)

## Description

Refreshes the in-repo Claude Code skill
(\`skills/react-native-executorch/\`) so it complies with the
[software-mansion-labs/skills contributing
schema](https://github.com/software-mansion-labs/skills#contributing)
and reflects the v0.9.0 API.

The previous skill was 661 lines in a single \`SKILL.md\` with eight
\`reference-*.md\` siblings; it was missing \`initExecutorch\`,
\`usePoseEstimation\`, \`usePrivacyFilter\`, and the new \`models\`
registry, and was still inlining raw URL constants.

What this PR does:

- Rewrites \`SKILL.md\` (138 lines) with proper YAML frontmatter
(\`name\`, \`description\` with trigger terms), a hook-selection
decision tree, critical rules, hook quick-reference table, and a
references table.
- Replaces the eight \`reference-*.md\` / \`core-utilities.md\` files
with four sibling files mirroring the upstream \`on-device-ai\` layout:
\`llm.md\`, \`vision.md\`, \`speech.md\`, \`setup.md\`. All four are
under 350 lines.
- Switches every code example from URL constants to the typed \`models\`
registry (e.g. \`models.llm.llama3_2_3b()\`, \`models.ocr.craft({
language: 'en' })\`, \`models.text_to_speech.kokoro.en_us.heart()\`).
- Adds coverage for v0.10.0 additions: \`initExecutorch\` +
\`ExpoResourceFetcher\` / \`BareResourceFetcher\` adapters,
\`useInstanceSegmentation\`, \`usePoseEstimation\`,
\`usePrivacyFilter\`.
- Removes \`skills/canary/\` (duplicate of the stable tree, differed
only by an \`internal: true\` metadata flag).

### Introduces a breaking change?

- [ ] Yes
- [x] No

### Type of change

- [ ] Bug fix (change which fixes an issue)
- [ ] New feature (change which adds functionality)
- [x] Documentation update (improves or adds clarity to existing
documentation)
- [ ] Other (chores, tests, code style improvements etc.)

### Tested on

- [ ] iOS
- [ ] Android

N/A — documentation-only change.

### Testing instructions

Skill files are consumed by Claude Code, not by the library runtime. To
sanity-check:

1. Verify the frontmatter on \`skills/react-native-executorch/SKILL.md\`
parses (\`name\`, \`description\`).
2. Confirm every reference link in \`SKILL.md\` resolves
(\`./references/llm.md\`, \`./references/vision.md\`,
\`./references/speech.md\`, \`./references/setup.md\`).
3. Spot-check a code example uses \`models.<category>.<name>()\` rather
than a raw URL constant.

### Screenshots

N/A

### Related issues

N/A

### Checklist

- [x] I have performed a self-review of my code
- [x] I have commented my code, particularly in hard-to-understand areas
- [x] I have updated the documentation accordingly
- [x] My changes generate no new warnings

### Additional notes

The reference-file layout (\`llm.md\` / \`vision.md\` / \`speech.md\` /
\`setup.md\`) mirrors
[software-mansion-labs/skills/.../on-device-ai](https://github.com/software-mansion-labs/skills/tree/main/skills/react-native-best-practices/references/on-device-ai),
so this skill can be lifted upstream as a refresh of that sub-skill in a
follow-up PR.

Author: msluszniak