test: cut fast-gate runtime ~104s->27s by removing redundant NNP model loads

[isayev]

Speeds up the default test gate (pytest -m "not slow") from ~104s to ~27s (3.85×) and eliminates the run-to-run variance.

Root cause (from pytest --durations)

~12 tests accounted for ~93 of the 104s — all loading the real AIMNet2 NNP (~4–10s each). The cache-semantics tests call ModelFactory.clear_cache(), evicting the shared model so redundant reloads landed on different tests each run (the 32–38s variance).

Changes (one decision per slow test)

• test_model_caching — fake AIMNet2Adapter via monkeypatch; caching is generic dict logic and the identity/size assertions stay meaningful (44s → 0.33s).
• test_model_adapter — the 5 AIMNet2 correctness tests share the session aimnet_model fixture (one load, not five); verified no in-place mutation of the shared adapter.
• test_model_factory — factory return-type tests reuse the session fixture (it survives clear_cache(), unlike a fresh create_model), so exactly one AIMNet2 load happens in the whole gate — this removes the variance.
• test_workflow / test_batchopt — stub create_model in tests that only exercise file-handling, bucketing, or already-faked optimization (run() returns before the model is used, or the optimization is monkeypatched).
• test_thermo_helpers — mark the two AIMNET Hessian-load checks @pytest.mark.slow (a separate ~9–10s model load, the single biggest remaining item; optional thermo path).

Coverage note

The fast gate now does exactly one real AIMNet2 load, kept deliberately for the adapter energy/forces correctness smoke test — so real-NNP numerical coverage still runs in CI (tests.yml runs -m "not slow"). Only the heaviest, lowest-value item (the Hessian load) moved to the slow suite.

Verification

• pytest -m "not slow": 629 passed, 8 skipped, 47 deselected — deterministic at 26.4 / 27.0 / 27.3s across three runs.
• Slow-marked Hessian tests still pass under -m slow.