feat(parse-label-filter): decide skip from which field the edit touched

[sydorovdmytro]

What

Adds optional body-changed and base-changed inputs so skip-edited is
decided from which field an edited event touched, instead of being inferred
from a body comparison that cannot see the difference.

Why

skip-edited inferred "the body before the edit" from
github.event.changes.body.from. That key is absent unless the body changed,
so every non-body edited event arrived with an empty previous body — which
parses as "the label-filter block was removed".

vcluster-pro's PR template ships a pre-filled block:

​```label-filter
none
​```

so for the default PR shape a title-only edit compares none against "",
reports a changed filter, and re-runs the suite for nothing. With the DEVOPS-1057
concurrency guard also in place, that run queues behind the in-flight one
rather than replacing it — the exact downgrade the guard and skip-edited are
paired to prevent. Found reviewing loft-sh/vcluster-pro#2156.

Comparing bodies cannot fix this, because the information is not in the bodies.

Approach

edited carrying	verdict	why
changes.title only	skip	title, labels, assignees cannot change a filter that lives in the body
changes.body	compare the blocks	unchanged from today
changes.base	never skip	a different base can mean a different thing under test

  body-changed: ${{ github.event.changes.body != null }}
  base-changed: ${{ github.event.changes.base != null }}

Why two inputs and not one

Both were tried as one and are wrong:

• body-changed alone reports a base retarget as skippable, since a retarget
  carries no body change either. A caller resolving an OSS branch from
  base_ref would then never test the new pairing. This was caught by Bugbot on
  the caller-side attempt in vcluster-pro#2156.
• changes.body.from || pull_request.body in the caller silences a body that
  genuinely changed from empty — a PR that just gained a label-filter block,
  which must run.

Both are now tests, so neither can be reintroduced.

Compatibility

Both inputs are optional and default to the legacy body comparison, so this needs
no coordinated rollout across the four vcluster-pro branches that call it. A
regression test pins the legacy misread, so the cost of omitting them is explicit
rather than folklore.

Only the exact strings "true"/"false" are acted on — a caller expression that
renders oddly falls back to comparing rather than producing a silent skip.

Tests

make test-parse-label-filter — 21 pass (8 new). shellcheck clean, docs
regenerated, make check-docs clean.

Follow-up

Wiring the two inputs into the callers (main, v0.34, v0.35, v0.36 of
vcluster-pro) once parse-label-filter/v1 is repointed. Not urgent — omitting
them is exactly today's behaviour.

Part of DEVOPS-1057.