ci: pin Strix Halo jobs to the native runners

[rominf]

Problem

"E2E tests (Strix Halo, Ubuntu)" has been intermittently failing on unrelated
PRs (#139, #140, #141), dying at its very first step with Permission denied
on mkdir. #143 passed only because it happened to be scheduled elsewhere.

Three self-hosted Strix Halo runners share the strix-halo label, and two of
them are Linux: the native Ubuntu host and a WSL host. The job's selector,
runs-on: [self-hosted, linux, strix-halo], matched both, so it landed on
whichever runner was free.

That job's setup hardcodes /home/ubuntu/actions-runner/... for HOME,
TMPDIR, and the pip cache — the native host's root partition is full and
that nvme is the only place with space. On the WSL host the runner executes
as a different user under a different home, so those mkdirs fail outright
and the job dies immediately. Failing runs report runner_name: strix-halo-wsl; the passing one reports strix-halo-ubuntu.

The nightly Strix job in nightly.yml copies the same hardcoded paths and the
same ambiguous selector, so it has the same latent failure — just less
visibly, since it is nightly and non-blocking.

Change

Add the native label to the runner selector for the Strix jobs, making
scheduling deterministic. The Windows jobs are already unambiguous in practice
(the WSL runner is labelled Linux, not Windows), but they are pinned too so
every Strix selector reads the same way and stays correct if the WSL host ever
gains a Windows label.

The comment above the Ubuntu job's env: block now notes that the hardcoded
paths are the reason for the pin, so the coupling is obvious to the next reader.

Notes

• Requires the native label on strix-halo-ubuntu and strix-halo-windows;
  this is already applied and live.
• Deliberately out of scope: making the hardcoded paths runner-agnostic, and
  adding any job for the WSL runner. Both are separate follow-ups.