Gate merge queue GPU CI by PR label (#9083)
The required `modal-torch-latest` workflow currently sends every
merge-queue entry with selected tests to a GPU backend. Many entries do
not need that GPU coverage, which consumes limited GPU capacity without
a per-entry opt-in.
With this PR merged, for master merge groups, the CI reads the exact
queue entry PR number from the event's head ref and check if the PR has
`needs-gpu-test `label.
Signed-off-by: Masahiro Tanaka <tanaka.masahiro@gmail.com>