Non-deterministic outcomes in forward simulation.

[pangtao22]

When comparing the gradients of the quasistatic dynamics using the python and cpp implementations, the tight tolerance can be satisfied about 9 out of 10 times. When the tolerance is not satisfied, the forward simulation results are also slightly off.

The current "solution" is to use a larger tolerance such as 1e-7, which shouldn't affect downstream tasks such as trajectory optimization. But if I really want to get to the bottom of this, I suspect the non-deterministic behavior is a result of drake's MathematicalProgram.