[Pallas/GPU interpreter] Relax test tolerances and use NumPy matmul for reference.

Tests failed in macOS CI due to small numerical differences between NumPy and JAX matrix multiplication outputs.

* Relax tolerances to 1e-6 in `gpu_pallas_interpret_tcgen05_race_detection_test.py`.
* In `gpu_pallas_interpret_test.py`, compute the expected array with `np.matmul` to match the interpreter's implementation and keep `assertArraysEqual`.

PiperOrigin-RevId: 972632869