Identify adequate stability of likelihood values

[itrharrison]

In e.g. #112 and #97 we have seen precomputed likelihood values change on O(0.01%) depending on the CAMB version. At the moment likelihood value stability is checked using np.isclose which has default settings on accuracy of rtol=1.e05, atol=1.e-8.

We should identify what rtol and atol values are actually appropriate for the SO likelihood codes.

[itrharrison]

TAC advice is that third decimal place accuracy should be adequate. I would propose that tests using allclose should have tolerances updated to reflect this, with the values being some package-wide parameter which can be changed if desired.