[Chore] Let the real-model tests run on another device (mps) #142

The test engine is fixed to CPU, so the mps half of #84's done criterion cannot be checked. Read GEMSTONE_TEST_DEVICE (default cpu) in api/tests/conftest.py and test_quantization.py.

Done when: GEMSTONE_TEST_DEVICE=mps uv run --extra torchnative pytest -m real_model builds the engine on mps; the default stays cpu.