Semi-breaking, in the sense that a variable that has been introduced since the last release has now been renamed.

Changes:

• Make test cases for solve-and-save-at not cover the product space of all configurations, but keep all but one parameter fixed while varying this parameter. Result: total tests reduce from >400 to <200, and test runtime goes from more than 550 seconds to approximately 200 seconds.
• Rename all "errorest" variables into "error" variables so that all error-estimator constructors fit into a single line in applications. Also, the "estimate" in error estimation should be implied.
• Merge the cached and non-cached versions of error estimators and distinguish them via a flag during construction. The rest of their code is too similar for two different objects to exist in the first place.
• Rename errorest_local_residual(_cached) into error_residual_std because the "local" is implied for error estimation, and because the estimator specifically uses the marginal standard deviation.
• Move the PDE benchmark to the basic examples because it doesn't have an external dependency
• Make normalisation in error estimation acknowledge higher-order information. Here is the before and after for the van-der-Pol and Pleiades benchmarks (which are second-order problems):

VdP Before:

VdP After:

Now, the fourth-order solver beats LSODA, and all solvers cover the ranges of precision more evenly.

Pleiades Before:

Pleiades After:

Now, the precision ranges are more evenly covered (similar to the non-probabilistic methods).