Currently the all_close tolerance is using the default of 1e-7, although our computations are in float64. We should decrease it to the same as in data tests and make it configurable per test/output. This might require generating smoother input data instead of fully random.