numerical issue for some realizations of atmosphere

I'm seeing large offsets in the photon x, y values, as well as the occasional nan, for some realizations of an atmosphere. The offsets can be very high, like 10 million pixels.

Below are example atmospheric parameters that produce the problem.

Using robust statistics on the output photons indicates that most photons land in a reasonable location, with a set of outliers

I can also provide a full example imsim config that produces this set of atm parameters

{'L0': [17.174188473465833,
        17.174188473465833,
        17.174188473465833,
        17.174188473465833,
        17.174188473465833,
        17.174188473465833],
 'altitude': [1.4302596307983027,
              5.348763607590827,
              11.208179050816556,
              18.24311343973453,
              24.467319796805825,
              30.175959028659584],
 'direction': [coord.Angle(5.023031335581778, coord.radians),
               coord.Angle(5.4138391122102645, coord.radians),
               coord.Angle(5.213667959962739, coord.radians),
               coord.Angle(5.043782770080258, coord.radians),
               coord.Angle(5.363918865113014, coord.radians),
               coord.Angle(5.075185823976301, coord.radians)],
 'r0_500': 0.08161434310901314,
 'r0_weights': [5.531557367962777e-13,
                4.3652382380659045e-14,
                1.326839454952542e-13,
                4.126186570683441e-14,
                2.344518516479255e-14,
                2.5838659853736628e-15],
 'rng': galsim.BaseDeviate('1778941255 1431167650 1282975060 ... 4184287197 2656677445 2301135708'),
 'screen_scale': 0.1,
 'screen_size': 800.0,
 'speed': array([10.94986395, 17.84175431, 37.29034581, 42.9768206 , 19.97968586,
       10.84201688])}

Start by reproducing the reported photon x, y offsets and occasional NaNs with the supplied atmospheric parameters, using the full imsim configuration mentioned in the issue. No source file or test is named, so trace the atmosphere-to-photon path from that reproduction; done means identifying and covering the failing realization without extreme offsets or NaN coordinates.