Make skills dual-compatible with Pi / the open-skills ecosystem

Summary

These skills are authored for Claude Code. The instruction bodies are already
agent-agnostic, but the packaging and docs assume Claude's discovery model, so
agents that use the open Agent Skills standard (e.g.
Pi) get a degraded experience: a misleading install flow and no catalog
manifest.

This PR makes the repo dual-compatible with additive changes only.
Nothing Claude-specific is removed — existing allowed-tools, compatibility
frontmatter, and the ~/.claude/skills symlink install flow are untouched.

What changed

• .claude-plugin/plugin.json (new) — declares the three skills
  (check-pr, greploop, cli-review) so the open skills CLI / catalogs can
  index the multi-skill repo.
• README — adds an "open-skills ecosystem" install path
  (npx skills add greptileai/skills) next to the existing Claude flow, plus a
  Pi usage section explaining progressive-disclosure activation (via the
  description field) and the explicit /skill:name command.
• metadata.triggers added to each SKILL.md as optional catalog hints for
  tooling that indexes skills. These are arbitrary metadata; agents that don't
  use them simply ignore them.

Why

• The Claude install instructions (~/.claude/skills + manual ln -s) don't
  apply to open-skills agents, which install directly from the repo and track
  installs in a skills-lock.json. The multi-skill layout already exposes each
  skill at <skill>/SKILL.md, so per-skill installs work without symlinks.
• A top-level plugin manifest is the convention these ecosystems use to index a
  repo that ships more than one skill.

Compatibility

• Claude Code: unaffected. No keys removed; symlink/slash-command flow intact.
• Open-skills / Pi: can now npx skills add greptileai/skills, and the docs
  correctly describe how activation works (description-driven, with /skill:name
  for explicit invocation).

Notes / accuracy

The README is careful not to overstate the mechanism: on Pi, skill
activation is driven by the description field (progressive disclosure), not by
metadata.triggers. The triggers are documented purely as optional indexing
hints.

Testing

• All metadata.triggers regexes compile cleanly (7 + 7 + 5).
• .claude-plugin/plugin.json is valid JSON and every referenced
  <skill>/SKILL.md path exists.
• Each skill remains self-contained and resolvable by
  npx skills add greptileai/skills/<skill>.

Out of scope

No changes to skill instructions, the references/ API docs, or the LICENSE.