test(preopt): skip GFN-FF tests when libxtb can't load under parallel load

[NCCU-Schultz-Lab]

Summary

Fixes flaky local failures of the GFN-FF (xtb) pre-optimization tests under pytest -n auto. Test-only change — preopt.py is untouched, and CI is unaffected.

Under heavy parallel load (~18 xdist workers), conda's libxtb shared library intermittently fails its first dlopen in whichever worker runs the GFN-FF relaxation tests, raising xtb C extension unimportable, cannot use C-API at run time. This turned test_relaxes_distorted_metal and test_trajectory_multiframe_ends_at_kept_geometry into flaky failures in the full suite only — they pass when the file is run alone, CI's pip-wheel xtb is unaffected, and the app itself already degrades to an honest no-op. So this is local test noise, not a product bug.

xtb_only now gates on _gfnff_runnable() — one real minimal GFN-FF run, cached per worker process:

• If libxtb won't load, the GFN-FF tests skip cleanly with an honest reason instead of failing.
• If it does load, the extension is primed (cached import), so the run-time flake can't recur for the rest of that worker's tests.
• A genuine GFN-FF regression still fails normally: the extension loads, the tests run, and their assertions catch it.

This also supersedes an earlier "broken C-API" hypothesis, which was a misread — import xtb.libxtb and a real GFN-FF run both succeed here; the cause is dlopen contention under parallel load, not a broken extension.

Test plan

• tests/test_preopt_gfnff.py alone — all pass.
• Full pytest -m "not network" under -n auto — 2573 passed, 29 skipped, 0 failed (previously 2 hard failures; the GFN-FF tests now skip cleanly in a worker where libxtb won't load).
• ruff + black clean.