CRPS for the t-distribution is not robust for large degrees of freedom and large magnitude of errors

[simon-hirsch]

The numerical robustness of the crps_t() suffers if the degrees of freedom are high and the errors are big.

and for dof = 100, but than we need to increase the size of the location error to provoke failing

Probably fall back to Gaussian for degrees of freedom larger than some constant, combined with a warning?

[sallen12]

Hi Simon, thanks a lot for pointing this out. The score depends on the cdf of the t distribution, which itself depends on the hypergeometric function. Scipy's implementation of the hypergeometric function was not robust to these cases, so I've adapted the implementation, which now seems to work better (see the PR at #96).

Let me know if you still have any issues with it.