Use RUNNER_TEMP for nvhpc tarball temp dir (ubuntu-26.04 tmpfs /tmp)

[djkees]

On ubuntu-26.04 (which ubuntu-latest moves to on 2026-10-19), /tmp is a ~7.8 GB tmpfs with a ~6.3 GB per-user quota, so extracting the nvhpc tarball (3.2 GB download for 20.9, extracted alongside) fails with tar: ... Cannot write: Disk quota exceeded. See actions/runner-images#14777.

• Put the tarball temp dir under RUNNER_TEMP (falling back to os.tmpdir()), matching src/installers/gfortran/debian.ts.
• Only the tarball path is affected (nvfortran 20.7, 20.9, and the apt-failure fallback); apt installs passed on 26.04 already.
• Rebuilt dist/index.js (one-line change).
• Adds an ubuntu-26.04 / nvfortran 20.9 job to ci-nvfortran.yml as a regression check.
• Adds a unit test that the tarball is downloaded and extracted under RUNNER_TEMP.
• Verified on 26.04 before this PR: unmodified action fails at extraction; with this change (or with TMPDIR set) 20.9 installs and the integration tests pass.
• The new nvfortran / 20.9 ubuntu-26.04 job passed in this PR's CI: the tarball was extracted under $RUNNER_TEMP (/home/runner/work/_temp/setup-fortran-nvhpc-*) with no "Disk quota exceeded", and the integration tests passed.