chore: remove MCP registry publishing

[macieju-opera]

The fork publishes to npm only, never to the MCP registry — the registry has no opera-devtools-mcp entry.

publish-to-mcp-registry-on-tag.yml came from upstream and triggers on chrome-devtools-mcp-v* tags, so it has never fired for an Opera release. It also has workflow_dispatch, so it could still be published by hand.

Removed:

• .github/workflows/publish-to-mcp-registry-on-tag.yml
• server.json — the registry manifest, its only consumer
• scripts/verify-server-json-version.ts and the verify-server-json-version npm script
• the "Verify server.json" step in pre-release.yml
• the server.json merge=opera-ours row in .gitattributes

A release now bumps package.json and package-lock.json instead of three files.

Intake safety

Deleting server.json is not free: upstream bumps it in every release commit, so it will conflict on each intake that spans an upstream release. Two things make that worse than it looks:

• merge=opera-ours does not apply to a delete. Git reports CONFLICT (modify/delete) and leaves upstream's copy in the tree, so a reflexive git add -A restores the file.
• Registering the path in docs/UPSTREAM.md made verify-upstream-seam accept it in any state, deleted or restored — so nothing caught the resurrection.

Second commit closes that: every path in the rename-or-delete table must now be absent from the working tree.

✗ 1 path(s) registered as renamed or deleted exist in the working tree:
    server.json
  An intake merge probably restored them: the merge driver does not apply to a
  delete, so git leaves upstream's copy in the tree. Fix: git rm <path>

Verified by restoring server.json and a renamed e2e test in turn — the check fails on each and passes on a clean tree. All nine paths already in that table satisfy it.

npm run gen produces no drift; format, typecheck and verify-upstream-seam pass locally.