test(integration): per-model examples + GPU real-checkpoint smoke tests

[drunkcoding]

What

Adds per-model runnable examples and real-checkpoint GPU smoke regression tests for the supported models, so the offload + generate() path has coverage per model before dev is promoted to main.

Examples (examples/)

deepseek_v2_chat_example.py (also fixes a dangling README reference), mixtral_example.py, qwen3_example.py, qwen3_5_example.py, glm_example.py, dbrx_example.py, jamba_example.py, olmoe_example.py, nllb_example.py.

Each follows the examples/readme_example.py template (MoE(ckpt, {offload_path, device_memory_ratio}) → tokenizer → generate), with per-model quirks: GLM-5.2 uses device_memory_ratio=0.5; NLLB uses the encoder-decoder translation path (src_lang, forced_bos_token_id); trust_remote_code set per model. GPT-OSS already has dflash_gpt_oss_example.py.

Regression tests (tests/python/integration/)

• test_model_smoke.py — parametrized over the 8 decoder MoEs (deepseek_v2, mixtral, qwen3, qwen3.5, gpt_oss, dbrx, jamba, olmoe): load real checkpoint → generate(max_new_tokens=16) → assert output grew + non-empty decode.
• test_nllb_smoke.py — dedicated NLLB-MoE translation smoke.
• GLM-5.2 already has test_glm_smoke.py.

Gating (won't touch CI): every test is @pytest.mark.gpu and guarded by its own MOE_<MODEL>_SMOKE=1 env var and a torch.cuda.is_available() skip. Heavy imports (torch/transformers/moe_infinity) are inside the tests. CI runs -m "not gpu" and doesn't collect tests/python/integration, so these skip cleanly and never block CI. Run locally with e.g. MOE_MIXTRAL_SMOKE=1 pytest tests/python/integration/test_model_smoke.py -k mixtral.

Verification

• py_compile all 11 files: pass
• ruff check: All checks passed
• pytest --collect-only: 9 tests collect, skip cleanly with no GPU/env
• Scope limited to examples/ + tests/python/integration/