remove interpolated specification of prior

[bkmi]

Evaluating the icdf and taking numerical derivatives doesn't work accurately, even with relative tolerance of 1e-4, for simple normal distributions (and not even some uniform distributions).

run the pytest tests/prior_test.py in the branch show_interpolation_doesnt_work.

@cweniger, I'm going to remove this prior specification option.

[bkmi]

Hmm... Further testing finds that a = dist.sample() a.cdf().icdf() != a even for pytorch normal distributions...

it works consistently with atol=1e-4 and rtol=1e-3.

[bkmi]

interpolated prior maps from v to u accurately using the same tolerance as in the last post.

however, maping from u to v is incredibly inaccurate. I did not find a tolerance which got this to work. I think we should remove the feature.

[bkmi]

Since the creator of this function @cweniger hasn't yet weighed in, I'm going to move this to long term. We can get it in the next release.

[cweniger]

This is no issue anymore in v0.4 since we don’t map on hypercubes.