Restore the release actor lookup the failure notice references

[ynaamane]

Summary

The release failure Slack notice reads needs.extract-release-actor.outputs.triggered_by,
but no job of that name exists in releaser.yml any more: #4982 removed
extract-release-actor together with the docs-website dispatch it fed, and left the
payload's reference behind. The expression is always undefined, so the notice falls
through to github.actor, which on release: published is the release GitHub App rather
than the person who started the release. actionlint flags it:

releaser.yml:364:428: property "extract-release-actor" is not defined in object type
{compute-build-flags: ...; image-build-and-push: ...; publish-helm: ...;
 release-binaries: ...; skills-publish: ...} [expression]

The producer side never went away: create-release-tag.yml still reads the
Release-Triggered-By trailer from the release commit and writes
<!-- Release-Triggered-By: X --> into the release notes "parsed by releaser.yml". This
restores the consumer.

• Re-add an extract-release-actor job that reads github.event.release.body through
  env: and extracts the username from the HTML comment (same pattern as the job removed
  in Unify docs-website release pipeline: new assets + retire dispatch chain #4982). permissions: {}, no token, no checkout, no ${{ }} inside run:.
• Keep only the first match (head -n 1): --generate-notes copies merged PR titles into
  the body verbatim, so a second Release-Triggered-By comment is reachable, and a two-line
  value would make the GITHUB_OUTPUT write malformed and fail the job.
• Add it to notify-release-failure's needs: so the output is addressable again.

When the comment is absent the output is empty, the step logs one line saying so, and the
payload's existing || github.actor fallback applies, as before.

Fixes #6265

Type of change

• Bug fix
• New feature
• Refactoring (no behavior change)
• Dependency update
• Documentation
• Other (describe):

Test plan

• Unit tests (task test)
• E2E tests (task test-e2e)
• Linting (task lint-fix)
• Manual testing (describe below)

The release workflow cannot run from a pull request, so the verification is static plus a
shell test of the extraction:

• actionlint .github/workflows/releaser.yml: the extract-release-actor property error
  is gone; the remaining output is byte-identical to main (two pre-existing shellcheck
  infos at lines 91 and 258, untouched here).
• zizmor --persona regular .github/workflows/releaser.yml: same findings as main, zero
  new (the new job binds the release body through env: and has no expression in run:).
• The new step's run: body executed locally under bash -e against sample release
  bodies: the comment present (yields the name), absent (empty value, the fallback log line,
  exactly one triggered_by= line written), an unrelated HTML comment first plus trailing
  whitespace (yields the name), and the comment twice with different names on one line and
  on two lines (yields exactly the first name and exactly one output line). Negative
  control: without head -n 1 the two-comment bodies write two lines. Also checked under
  set -o pipefail: a no-match body exits 1 and a two-match body exits 141 from head
  closing the pipe, which is why || true stays.
• Removing the new needs: entry brings the actionlint error back; removing the job while
  keeping the needs: entry produces the unknown-job error. Both restored.

Does this introduce a user-facing change?

No.