feat: Auto-bump astronomer-data plugin version on content changes

[kaxil]

Summary

The astronomer-data plugin has shipped from .claude-plugin/plugin.json with "version": "0.1.0" since the manifest was first created, and the string never moved. Claude Code resolves a plugin's version from plugin.json first, and skips an update whenever the resolved version matches what a user already has. So every user — including everyone on the official Anthropic marketplace, where we're listed as astronomer-data-agents, data, and data-engineering, all with "version": null (which falls through to this file) — has been frozen on whatever they first installed. New skills never reached them.

This PR does two things:

1. Catches the version up to 0.2.0 so the next time our pinned sha is refreshed, the accumulated skill changes since 0.1.0 actually get delivered.
2. Adds a workflow that auto-increments the patch version on every push to main that touches skills/** or the marketplace manifest — so we can't silently re-freeze.

Design rationale

• Why automate instead of bumping by hand? Manual bumping is exactly what failed — the field sat at 0.1.0 across dozens of skill changes. The workflow makes the version advance a side effect of merging content.
• Why patch-only auto-bump? Monotonic increase is all that's needed for update detection to fire. Deliberate minor/major bumps are still done by editing plugin.json in the PR; the workflow detects that the push already changed the version and skips its own bump, so it never fights a manual decision.
• Why exclude plugin.json from the trigger paths? The bump commit only touches plugin.json; excluding it from paths stops the workflow from retriggering itself.
• Format-preserving edit. The bump replaces only the version value via sed, leaving indentation and the trailing comma intact, so the manifest stays clean in diffs.
• This mirrors how peer data-platform plugins handle it — Databricks (databricks-agent-skills) sits at 0.2.9 and Snowflake (snowflake-ai-kit) at 3.2.2; both bump on release rather than freezing.

Gotchas / tradeoffs

• Direct push to main. The workflow commits the bump back to main using the default GITHUB_TOKEN. If main is protected against pushes from Actions, either grant github-actions[bot] a bypass, or swap the final step for a PR (e.g. peter-evans/create-pull-request) — flagged inline in the workflow.
• Fork PRs that change skills won't auto-bump on their own merge (fork-token limitations); the next intra-repo push to main will catch the version up.
• Every skill-touching merge now advances the patch version. That's intended — each change is meant to reach users — but it does mean a steady stream of chore: Bump ... commits.

Follow-up (not in this PR)

The bigger reach lever is that the official Anthropic marketplace pins our sha and only re-points it periodically, and we currently appear under three names there (astronomer-data-agents, data, data-engineering). Worth a separate cleanup/refresh conversation.